3.1. The engagement system

The engagement system [6,10,11], part of Martin and White’s [6] appraisal framework, involves discourse semantic devices that enable the authorial voice to position itself relative to and engage with other voices and alternative perspectives within a communication context. Grounded in the notion of dialogism [31], the engagement system emphasises heteroglossic markers in discourse. These markers show how writers/speakers make proclamations and negations, acknowledge alternative voices, entertain alternative perspectives, and distance themselves from particular viewpoints. These theoretical categories emphasise not only linguistic markers but also how writers negotiate relationships with different audiences.

In the engagement system, the Proclaim category encompasses devices used to align with textual voices, which are categorised into three subcategories: Pronounce, Concur, and Endorse [6,11]. Pronounce involves explicit intervention in the text, where speakers or writers openly express their perspectives. Examples include “we affirmed that”, “we are committed to”, and “our team contends that”. Concur is used to express alignment with textual voices by presenting them as uncontentious and universally accepted. Phrases such as “of course”, “it is common sense that”, and “clearly” imply a shared understanding between speakers/writers and the putative audience, enhancing a sense of agreement. Endorse ascribes credibility to propositions from external sources, positioning the speaker or writer as an advocate of these viewpoints. The strategy is expressed by verbs such as “found” and “showed”, as in “the study participants found that” and “a May 2020 study showed that”(for similar classification, see [32,33] among others).

Disclaim encompasses devices that serve to contradict or challenge textual voices. They are categorised into two subcategories: Deny and Counter [6,11]. Deny is used to introduce propositions that are then explicitly rejected or dismissed. The pattern is exemplified in the negation “no” in “no team is responsible for”. Counter acknowledges an initial proposition and then introduces a contrasting view, using conjunctions such as “although”, “but”, and “however”.

Expansion that involves Entertain, Acknowledge, and Distance allows speakers or writers to entertain other propositions and attribute external voices. As defined in the theory [6,11], Entertain presents the authorial voice as one of many possible propositions, acknowledging the validity of alternative perspectives. This subcategory includes likelihood devices, such as modal auxiliaries (“may” and “could”), modal adjuncts (e.g., “perhaps”), and modal attributes (e.g., “it is likely that…”). It also involves evidential expressions (e.g., “we believe” and “it seems”) and appearance-based postulations (e.g., “it appears”) (for similar classification, see [8,34,35] among others). Acknowledge presents attribution of external voices without explicit endorsement, using reporting verbs such as “say”, “state”, and “announce”. Distance, on the other hand, separates the authorial voice from attributed material, using verbs such as “claim” and scare quotes.

The engagement system has been widely used to examine rhetorical appeals in academic and legal discourse. These studies have enriched the theoretical foundation by revealing the pragmatic functions of engagement strategies in professional discourse. In academic writing, Endorse appears in citations for authoritative assertions [32,33]. Counter enables detailed discussions and presents alternative propositions [35], while its combination with Deny conveys caution in findings [34,36,37]. Entertain and Deny jointly mitigate extreme statements in social science literature [34,36,38]. Acknowledge makes subtle references to prior research [38]. Conversely, engagement is widely used in legal discourse of various types [39,40]. Judges use Pronounce to grant appeals, outline case details, and indicate judicial agreement. They also use Endorse to align with precedents and Deny to refute arguments and dismiss appeals [39,40].

Many studies have applied this theory to corporate reports, illuminating the distinctive pragmatic meanings of these persuasive tactics within this professional genre. Proclaim and Disclaim are used in organisations’ responses to government quality audits. Combining with Engage, they enhance the institution’s self-presentation [8]. Within Proclaim, Pronounce is used in CSR reports to articulate the determination to achieve its goals [9,20,26]. Endorse in annual reports projects a confident and commanding corporate image [41]. Additionally, scholars have increasingly emphasised Entertain in annual and CSR reports. Likelihood devices in annual reports show companies’ rationality and decision-making ability [41], while evidential expressions in CSR reports portray organisations as ethical and socially responsible entities [20,26]. In CEO statements of CSR reports, evidential formulations express corporate missions and initiatives [9,42], while modal devices convey organisational capabilities and benevolence [9]. Despite such insights, existing research tends to foreground the pragmatic meaning of engagement markers while under-exploring who is engaged and what information is conveyed systematically. These dimensions arguably warrant equal consideration, as they potentially shape persuasive meaning construction in corporate reports [25].

We applied this framework to AI ethics statements for two reasons. It encompasses a broader range of stakeholders than other linguistic theories on written discourse. In contrast to metadiscourse theory [43,44], which primarily examines how writers cultivate rapport with readers, Martin and White’s engagement system shows how writers position themselves in relation to multiple stakeholders. The framework captures not only who is engaged but also which perspectives are aligned with different audiences. This dual focus enables us to map particular AI ethics performances onto their intended stakeholder groups. On the other hand, compared with conversational pragmatic theories, this framework better fits discourse-level analysis. Whereas speech-act approaches model the illocutionary force of individual utterances (e.g., directives) [45,46], the engagement system examines dialogic resources across extended texts. Since AI ethics statements are carefully crafted documents rather than spontaneous interactions, a discourse-level perspective yields more relevant insights. The theory thus allows us to examine how companies strategically shape their ethical identities in extended texts through stance management, negotiation of heteroglossic space, and reader alignment.

Drawing on this framework, this paper proposes a two-dimensional engagement typology that systematically specifies the targets addressed and the parties involved. We examine how engagement strategies prioritise various dimensions of corporate ethical performance, including overall strategies (Strategic target), external voices and/or actions (Responsive target), and operational practices (Operational target). This analysis reveals which aspects of corporate performance AI ethics statements strategically emphasise. Additionally, we further show how AI companies tailor engagement strategies to two stakeholder groups, underscoring the value of integrating business-theory concepts into corporate discourse studies. By combining these two analytical dimensions within engagement analysis, this project advances understanding of how AI companies adapt their engagement strategies in broader CSC contexts.

3.2. Concepts in stakeholder theory

This paper uses the concepts of stakeholders and the classification of primary and secondary stakeholders from stakeholder theory [5,12] to identify relevant parties in the statements.

Stakeholders are “groups and individuals who can affect an organisation” when a company decides to achieve specific aims [12, p.48]. This broad definition acknowledges the diverse operational environment and highlights CSR’s role in enhancing business performance, including AI ethics compliance [47]. Strategic management scholars use these concepts to evaluate CSR performance against stakeholder expectations [5,48].

The present study adopted Clarkson’s [5] categorisation of primary and secondary stakeholders based on their risk to company actions and viability. Primary stakeholders are critical to a company’s survival and success [5]. They engage in market activities, provide financial resources, and participate in important social interactions [49]. Conversely, secondary stakeholders are individuals and communities who “are not engaged in transactions with the company and are not essential for its survival” [5, p.107]. They exist in the corporation’s societal sphere, including media, academic scholars, and non-governmental organisations [49]. Such classification enables researchers to examine organisational strategy tailoring for different groups, based on their importance to the corporation’s success [48].

This paper adopts Clarkson’s primary and secondary stakeholder framework because it aligns precisely with how firms structure their CSR/ESG communication [50]. While frameworks such as Freeman’s [51] view stakeholders as non-hierarchical, he recognises that companies prioritise stakeholders based on their impact on organisational survival. This essentiality-based hierarchy drives communication strategies. Empirical evidence has confirmed this pattern: companies employ differentiated approaches in their CSR/ESG reports when addressing primary versus secondary stakeholders [4,48,52]. Linguistic analyses of this discourse genre also reveal distinct discourse features for specific audiences [20,23,25]. Since AI ethics statements serve as a form of CSR/ESG communication, they should exhibit similar differentiation in their linguistic engagement tactics. Clarkson’s classification thus offers the ideal analytical tool for our analysis. By distinguishing between the two stakeholder groups, this framework enables systematic analysis of how AI companies engage different groups through specific linguistic choices.

4.1. Data collection and data

A data collection of publicly published AI ethics statements from 50 leading international AI developers and adopters that have played a major role in shaping the global AI landscape. Because AI development and adoption were especially active in the U.S., UK, Japan, and China, our corpus focuses on 50 AI companies. Of these, 48% (n = 24) are headquartered in the U.S., 12% (6) in the UK, and 12% (6) in China. The remaining firms are based in Japan (3), Sweden (3), Germany (2), South Korea (2), the Netherlands (1), Norway (1), Canada (1) and Finland (1). This distribution reflects real-world language use while balancing regional representation. Statements were obtained from the official homepages of the chosen firms between 1 May 2024 and 30 June 2024. Data collection consists of two steps: 1) identification of AI companies and 2) corpus compilation.

4.1.1. Identification of AI companies.

Company identification began with top ranks from Sequoia’s AI 50 2023 [53] and TIME100AI 2023 [54]. As shown in Fig 1 above, the two prominent industry lists were consolidated to identify companies actively engaged in AI development. After removing duplicate entries, this process yielded 56 distinct AI development companies (e.g., OpenAI, Anthropic, and Meta). The geographical distribution shows a pronounced concentration in specific regions: 39 companies were headquartered in the U.S., while the remaining 17 were primarily located in the UK, China, and Japan. This distribution pattern underscores the geographic concentration of AI development activities within these key markets.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Workflow for selecting AI companies.

https://doi.org/10.1371/journal.pone.0340796.g001

A comprehensive examination of each company’s official website was subsequently conducted to identify publicly available documents addressing AI ethics considerations. The search encompassed various document types, including corporate policies, annual reports, press releases, and blog posts. These materials were predominantly categorised under the sections labelled “company stories”, “corporate reports”, and “policy statements”. Through keyword searches using terms such as “data security” and related ethical frameworks (as shown below), 18 companies were identified as having published English-language documentation specifically addressing AI ethics concerns, including OpenAI, Meta, and Google. In contrast, we found no such AI ethics statements for the remaining 38 companies, including Hugging Face, Ironclad, and PathAI.

The majority of companies maintaining public AI ethics documentation were based in the U.S. This geographic skew in the initial sample toward American AI companies necessitated expanding the data collection methodology to ensure broader international representation. To address this limitation, ChatGPT 4o was employed using a prompt (see S1 Appendix) to generate an additional list of globally recognised AI companies. This prompt specifically requested identification of leading AI development companies, such as OpenAI, alongside major AI application companies, such as Mastercard, while excluding the previously identified 56 firms. This process generated 62 additional companies for evaluation.

Identical search procedures were applied to examine the corporate websites of these newly identified firms, specifically targeting English-language corporate disclosures related to AI ethics practices and policies. Such an expanded search yielded 32 companies with relevant documentation. Two independent researchers reviewed each company’s homepage to assess its suitability, focusing on its use and development of AI and any publicly available documents on AI ethics. As a result, these firms were subsequently incorporated into the data collection, bringing the total sample of 50 AI companies.

4.2. Analytical procedures

The data analysis comprised five steps: (1) identifying primary and secondary stakeholders, (2) exploring engagement strategies, (3) identifying engagement targets, (4) performing intercoder reliability tests, and (5) identifying statistical differences (Fig 1). Coding was conducted during the first three steps, encompassing eight stakeholder types (Step 1), seven engagement markers (Step 2), and three engagement targets (Step 3), for a total of 18 categories. Targeted engagement—the combination of engagement strategies and engagement targets—was then counted at the sentence level and aggregated by stakeholder classifications. All statistical comparisons were performed at the corpus level without weighting by firm. Descriptive statistics on firm-level publication counts were provided in the S1 Appendix.

4.2.2. Identification of engagement strategies.

The second step involved annotating engagement strategies in the sentences addressing participants. Based on the engagement system [6,10,11], seven strategies were discerned at the word and phrase level. Table 1 summarises the engagement taxonomy used in this paper.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Taxonomy of engagement.

https://doi.org/10.1371/journal.pone.0340796.t001

The main challenge in annotation is identifying simultaneous engagement markers. When multiple engagement markers coexisted in a sentence, each was annotated separately to maintain analytical precision (see Example 2).

1. (2). If a product does not perform as well for some people, perhaps because of inherent skin tone biases, that might be a product liability issue [66].

Example 2 uses Operational Deny (“does not”) and two instances of Operational Entertain (“perhaps” and “might”). They acknowledge the possibility that AI systems could discriminate against certain interest groups, particularly those expressing interest in innovations and exhibiting diverse skin tones.

6.1. Primary stakeholders: Operational Deny

As previously stated, Operational Deny refers to the negation of certain operational benefits by AI companies to show adherence to ethics. It was significantly more prevalent (p < 0.0001) in sentences targeting primary stakeholders. In these sentences, Operational Deny primarily addressed customers and third-party developers (154 instances, 93.33% of 165). Such expressions rejected others’ propositions [34,37] and demonstrated data security commitments to primary stakeholders. Common examples included “no”, “not”, “will not”, and “cannot”. For instance, the sentence “Huawei will not store users’ product search records” [75] assures customers that their personal data remains private.

Operational Deny co-occurred with other targeted strategies in one sentence. The most common combination paired Operational Entertain and Operational Deny (78 instances, 47.27% of 165 instances) through expressions such as “may…not”, “may…deny” and “not…would”. As shown in Fig 2, they appeared predominantly when engaging with customers and third-party developers, accounting for 71 instances (73.96% of the pair’s use across stakeholder groups). The purpose was to allow AI companies to communicate privacy policies without making absolute claims (cf. [33,50]. Thus, users were informed of the options to refuse data collection, as in Example 4.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Distribution of frequent combinations involving Operational Deny across stakeholder groups.

https://doi.org/10.1371/journal.pone.0340796.g002

1. (4). If you do not have an account with us, we may ask you to provide additional Personal Information and proof of residency for verification [76].

Example 4 shows OpenAI’s use of Operational Deny and Operational Entertain when sending a conditional disclaimer to unregistered users. The disclaimer begins with a condition that excludes registered users. This Operational Deny clarifies the scope of the disclaimer and ensures it is intended only for unregistered users. OpenAI then uses Operational Entertain (“may”) to present the option to request additional personal information. Such a modal verb simulates these customers while ensuring the firm’s full discretion over any actual interaction.

The less common combination of Operational Counter, Operational Deny, and optional Operational Entertain (19 instances, 11.52% in 165 instances) appeared in the sub-corpus. As in Fig 2, they primarily engaged with customers and third-party developers (16 instances, 80% in the pattern use across stakeholder groups) with patterns of “but…not”, “only…not”, and “however … cannot… may”. The combination facilitated the presentation of disclaimers, as in (5), and helped the firms to acknowledge efforts and identify their limitations [30].

1. (5). However, despite our safeguards and efforts to secure your information, no electronic transmission over the Internet or information storage technology can be guaranteed to be 100% secure, so we cannot promise or guarantee that hackers, cybercriminals, or other unauthorised third parties will not be able to defeat our security and improperly collect, access, steal, or modify your information [77].

In (5), Stability AI’s disclaimer employs engagement strategies to address data storage risks. The company uses Operational Counter, characterised by the word “however”, to raise the reader’s expectations of its statement [38]. Another Operational Counter (“despite”) acknowledges efforts in safeguarding information before presenting limitations [6,10,11]. The use demonstrates the firm’s commitment to data protection, potentially helping maintain user confidence and attract new customers. Subsequently, Operational Deny, marked by “no”, “cannot”, and “will not”, refutes assumptions about perfect safety and reinforces the limitations of security measures. This targeted engagement manages user expectations and limits liability for potential data breaches. A transparent and responsible approach to data security communication, as recommended by GRI [17], is thus deliberately adopted. This allows the firm to strategically balance transparency requirements with reputation management, addressing stakeholder needs while protecting corporate interests.

6.2. Primary stakeholders: Operational entertain

Operational Entertain is used by AI companies to present potential ethical practices in data management, product development, and service implementation. This tactic appeared significantly more often (p < 0.0001) in the primary stakeholder sub-corpus, primarily addressing customers and third-party developers (486 instances, 84.38% of 576 instances). The evidential markers “we believe” and “we recommend” negotiated companies’ views among alternatives. The epistemic modal “can” conveyed companies’ capabilities in responsible AI practices. These discourse choices reflect corporate efforts to present themselves as socially responsible entities [9,23]. The likelihood device “may” presented potential benefits to primary stakeholders while fostering mutual trust and long-term business success.

In the primary stakeholder sub-corpus, Operational Entertain, although relatively infrequent, occurred with Operational Pronounce (53 instances, 9.2% in 576 instances). As shown in Fig 3, the combination was primarily used to engage customers and third-party developers (53 cases, representing 94.64% of all instances where this pattern was applied). Typical expressions included “we know... we will continue to”, “it is important to note that… you may”, and “we believe you must”. This pair of targeted engagement strategies involved primary stakeholders by demonstrating the potential benefits of companies’ AI ethics management practices, as in (6).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Distribution of the combination Operational Entertain and Operational Pronounce across stakeholder groups.

https://doi.org/10.1371/journal.pone.0340796.g003

1. (6). We know people want more transparency and control over their data, and we will continue to develop new and improved privacy tools and controls to meet those expectations [78].

At the beginning of (6), Operational Entertain, signalled by the evidential maker “we know”, shows Meta’s efforts to recognise and likely empathise with customer expectations for data control. This is followed by Operational Pronounce (“we will continue to”), which explicitly shows Meta’s pledge to fulfil these expectations. Combining these markers indicates that the company is aware of key stakeholders’ concerns and is committed to future actions. Meeting users’ information needs is not only consistent with established practices in IFRS S1 [18] but also indicates that Meta prioritises customer requirements.

6.3. Primary stakeholders: Responsive acknowledge

Responsive Acknowledge is used by AI companies to attribute suggestions to external sources and acknowledge their existence without explicitly endorsing or distancing themselves from them. This strategy was significantly more common (p < 0.01) in communicating with primary stakeholders. It was utilised to report on their AI ethics initiatives by citing senior management’s views (66 instances, 66.67% of 99 instances). Adopting decision-makers’ voices increased disclosure credibility [25] by revealing how executives approach AI ethics. Additionally, the use of verbs such as “said”, “pointed out”, and “suggested” conveyed a neutral stance towards management subjectivity, strengthening the objectivity of reporting on corporate social practices. This approach aligns with international criteria for impartial disclosure of sustainability management [18] and meets stakeholders’ information needs by acknowledging uncertainty in plans.

The combination of Responsive Acknowledge and Strategic Pronounce (18 instances) appeared in 18.18% of 99 sentences addressed to primary stakeholders. As in Fig 4, they were used particularly to engage with corporate management (11 instances, 61.11% of all cases in which this combination was applied). Expressions such as “pointed out … we should” informed about AI ethics plans while maintaining neutrality. They aligned management commitments with corporate strategic AI ethics goals, enhancing organisational neutrality, as in (7).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Distribution of the combination Responsive Acknowledge and Strategic Pronounce across stakeholder groups.

https://doi.org/10.1371/journal.pone.0340796.g004

1. (7). Besides, Mr Tian further pointed out that, in the post-pandemic era, we should promote the accelerated application of digital technologies and unleash and enjoy digital benefits. At the same time, we should also give full play to the advantages of technology and innovation in closing the “resource gap” more quickly and creating shared benefits, bringing the benefits of the digital economy to people across all races, generations, urban and rural areas, and regions, enhancing social justice and fairness [79].

In (7), SenseTime uses Mr Tian, head of its intelligence research centre, to articulate its vision for socially responsible digital technology advancement. With the Responsive Acknowledge marker, “pointed out”, the AI solutions provider presents Mr Tian’s views while maintaining neutrality. The statement features two instances of Strategic Pronounce, marked by the repeated use of “we should”. These obligation modals indicate management’s strategic focus on both digital technologies and equity. Such dual commitments balance technological progress and equitable distribution of benefits, while emphasising the inclusivity of diverse populations. By attributing these strategic commitments to a high-level executive through Responsive Acknowledge, SenseTime objectively highlights its pursuit of fairness [1] while deliberately leaving the actual technological feasibility ambiguous.

6.4. Secondary stakeholders: Strategic pronounce

Strategic Pronounce is a tactic to overtly express AI companies’ business plans. The quantitative analysis indicated that Strategic Pronounce was significantly more common (p < 0.0001) in communication with secondary stakeholders. In 252 instances, 69.84% targeted society (176 cases), while 12.3% (31 cases) addressed interest groups (e.g., tech-savvy groups and industry associations). Corporate determination was signalled through willingness adjectives “committed” and “dedicated”, and the modal “will” [9,20,24]. Phrases such as “we are committed” and “we will continue” conveyed a sense of resolve. Goal-oriented expressions such as “our goal is” and “we strive to” emphasised a commitment to implementing AI ethics. These linguistic choices illustrate the firms’ strong confidence in their ethical convictions and convey an image of socially responsible organisations.

Strategic Pronounce frequently co-occurred with other targeted engagement strategies in one sentence. It combined Strategic Entertain in 18 instances (7.14% of 252 instances), allowing AI companies to express potential benefits while maintaining flexibility, as in (8). As in Fig 5, this combination was particularly used to engage with the whole society (15 instances, 53.57% of the use of this combination across stakeholder groups). They appeared in expressions, such as “we will continue to … could”, “our vision is … may”, and “Importantly, we believe…”. Through these formulations, the firms demonstrate their awareness of social welfare and academic contributions, thereby strategically building their image of social responsibility [8]. However, Strategic Entertain undermines the very promises articulated through Pronounce, enabling firms to evade concrete obligations [45].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Distribution of frequent combinations involving Strategic Pronounce across stakeholder groups.

https://doi.org/10.1371/journal.pone.0340796.g005

1. (8). We will continue to work on alternative and more inclusive measures that could be useful in the development of our products in collaboration with scientific and medical experts, as well as groups working with communities of color [80].

In (8), Google’s strategic direction in developing AI ethics is expressed using Strategic Pronounce in the opening sentence (“We will continue to”). This assertive statement emphasises the dedication to promoting inclusivity and suggests broader social benefits by engaging with specific social groups. The firm then uses the Strategic Entertain word “could” to discuss the potential benefits for both Google and its stakeholders, legitimising its commitment [26]. The low epistemic certainty of the word distance Google from potential failure to deliver on these social objectives. The firm further shows its dedication to inclusivity by highlighting its collaboration with scientific and medical experts and groups representing communities of colour. To summarise, this practice shows an intertextual connection with the consideration of accessibility embodied in the U.S. National Artificial Intelligence Initiative Act of 2020 [30].While aligning with regulatory frameworks, this measured language preserves necessary flexibility given the evolving nature of AI development.

Furthermore, Strategic Pronounce co-occurred with Strategic Deny in 12 instances (4.76% of 252 instances), while in 2.23% of cases, Strategic Counter complemented this combination. Fig 5 shows that this co-occurrence appeared mainly in the engagement with the whole society (12 cases, 63.16% of the pattern use). The typical expressions included “In particular … we will not”, “shall be … does not”, and “we should … not”. Unlike the previous encounter, it allowed the AI corporations to define strategic positioning by enforcing their preferred corporate missions and rejecting alternatives, as in (9).

1. (9). To develop trustworthy AI, it’s key to consider not just what data is legally available to use, but what data is socially responsible to use [81].

Example 9 illustrates how Nvidia conceptualises its vision of trustworthy AI through the lens of engagement. The company uses Strategic Pronounce (“it is key to”) and Strategic Deny (“not just”) to reject the notion that legal compliance alone is sufficient for ethical AI development. This combination emphasises Nvidia’s stance against a purely legalistic approach to AI ethics, which focuses on the rule of law. Nvidia then uses Strategic Counter, characterised by the conjunction “but”, to present an alternative perspective. This counter-position emphasises the importance of social responsibility in data use, extending the concept of trustworthy AI beyond legal boundaries. Through this framing, Nvidia not only defines what trustworthy AI ought to look like but also advances its strategic positioning by asserting a preferred mission and refuting legislative alternatives.

6.5. Secondary stakeholders: Strategic Counter

Strategic Counter is used by AI companies to introduce and supplant contrary claims to their corporate strategic plans regarding AI ethics. It was significantly more common (p < 0.0001) in the sentences directed at secondary stakeholders, primarily society (71 cases, 82.56% of 86 instances) and academic institutes and researchers (9 cases, 10.47%). Strategic Counter typically employed conjunctions and adverbs such as “however”, “but”, and “while” [6] to confirm and then refute oversimplified or ethically problematic ideas about AI implementation. This targeted engagement maintains reader solidarity [6] while emphasising how AI firms commit to ethical AI development, as in Examples 10 and 11.

Strategic Counter frequently occurred with Strategic Entertain in 24 instances (27.91% in 86 cases) and Strategic Deny in 12 instances (13.95%). Strategic Counter and Strategic Entertain combined to contradict prevailing views on AI ethics and then promote the companies’ potential for responsible AI solutions (see Example 10). As shown in Fig 6, the combination appeared mainly in engagement with society (7 instances, 18.42% in the pattern’s use across stakeholder groups) and with research institutes and scholars (14 instances, 36.84%). Typical expressions involved “but…can”, “although…we believe”, and “even so…can”. Rather than hedging against readers’ expectations in other professional discourse genres [34], Strategic Entertain has weakened the assertiveness of corporate solutions.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Distribution of frequent combinations involving Strategic Counter across stakeholder groups.

https://doi.org/10.1371/journal.pone.0340796.g006

1. [10]. Catastrophic risks are a possible or even plausible outcome of advanced AI development. Counteracting this requires a substantial scientific and engineering effort, but with enough focused work we can achieve it [82].

Anthropic’s example shows the use of Strategic Counter and Strategic Entertain in risk management. It starts with recognising potential risks and building credibility through transparency. “A substantial scientific and engineering effort” connotes the necessary collaboration with academia, indicating the complexity of the challenge. This engagement enhances the scientific credibility of AI development and underscores the challenges of achieving it. Having established the problem, Anthropic employs a Strategic Counter (“but”) to pivot from a problem to a solution, reframing a risk as a manageable business opportunity [27]. The phrase “with enough focused work” outlines a concrete collaboration with scientific scholars and reinforces credibility. Then, Strategic Entertain “can” in “we can achieve it” balances confidence with measured optimism.

Moreover, the combination of Strategic Counter and Strategic Deny created a more decisive debating tone by explicitly refuting a claim and asserting its alternative. It commonly employed expressions such as “not…but”, “don’t… but”, and “doesn’t… but” when engaging with society (10 cases, representing 50% of all instances where this pattern was used, as shown in Fig 6). While this combination is common in professional communication [8,34,38], it was uniquely used in AI ethics statements to firmly refute unethical propositions and clearly articulate the company’s ethical stance, as in (11).

1. (11). Not everyone speaks perfectly, but everyone deserves the same level of service [83].

Example 11 from Ericsson’s white paper presents its vision of promoting fairness and inclusivity in AI model development (not reported here). The statement begins with Strategic Deny (“not”) to reject a discriminatory presupposition and resonates with social expectations. The subsequent Strategic Counter (“but”) elicits Ericsson’s promise to ensure equal service for all, foregrounding the organisational value of social justice. Thus, the expression “not...but” acknowledges AI developmental challenges while highlighting Ericsson’s emphasis on fairness and non-discrimination. This approach not only reflects UNESCO’s [1] social justice mandates but also addresses the public’s emotional needs and enhances trustworthiness within the community [21].

7.1. Findings

Our analysis shows a significantly higher frequency (p < 0.001) of targeted engagement with primary stakeholders, in particular customers, third-party developers, and corporate management. AI companies primarily employed Operation-targeted tactics, including Operational Deny, Operational Entertain and Operational Counter, to communicate service updates and security protocols, addressing data security concerns. These strategies frequently occurred in combinations to involve primary stakeholders. Operational Entertain was frequently paired with Operational Deny and Operational Pronounce. Another common combination involved Operational Counter, Operational Deny, and occasionally Operational Entertain. In addition, companies also employed Responsive Acknowledge to convey managerial communication and Strategic Endorse to show compliance with local regulatory authorities. The most frequent combination was Responsive Acknowledge and Strategic Pronounce.

AI companies shifted to Strategy-targeted engagement for secondary stakeholders, specifically society, interest groups, and academia. Strategic Pronounce, Strategic Counter, and Strategic Deny were employed to negotiate AI ethics commitments and overall development blueprints. Strategic Pronounce was commonly paired with Strategic Entertain and Deny, while Strategic Counter was frequently paired with these two strategies to address the needs of secondary stakeholders.

7.2. Theoretical implications

These stratified discourse strategies for primary and secondary stakeholders create a linguistic engagement framework that large AI firms have employed in their communications about social responsibility and service orientation. This framework aligns with previous linguistic studies on CSC [22–25,42] that show companies adjust their discourse strategies to audience preferences across cultures. Building on that insight, our framework identifies how world-leading technology firms communicate about diverse stakeholder concerns regarding AI performance. The interplay of linguistic and business theories underlying this framework shows how they present themselves ethically in their public discourse.

Regarding engagement with primary stakeholders, this paper identified both the strategies found in existing CSC studies and more specific patterns. Previous studies have shown that epistemic and evidential markers frequently appear in CSR/ESG reports to demonstrate the capabilities and future practices [20,23,24,41]. Building on this scholarship, we found that the combination of Operational Entertain (characterised by epistemic and evidential markers) and Operational Pronounce enabled firms to frame potential customer benefits with built-in hedges (section 6.2). They help construct provisional narratives that accommodate both optimistic projections and emerging technological limitations. Moreover, while Bondi and Yu [25] argued that managerial remarks in CSR reports increase credibility, we observed a more nuanced pattern. The managerial voice appeared in Responsive Acknowledge and Strategic Pronounce to maintain distance from direct company voice (section 6.3). This combination suggests caution in dealing with uncertain technologies, given the current climate of public concern, while maintaining strategic manoeuvrability.

Additionally, corporate AI ethics statements also employed patterns from scholarly communication, particularly Endorse [29,30,40] and Deny [34,37,73]. For instance, the coexistence of Operational Deny and Operational Entertain created a discursive space in which companies acknowledged their limitations while maintaining openness to multiple perspectives. Similarly, the combination of Operational Counter and Deny employed an “anticipate-and-refute” strategy to discuss sensitive topics with primary stakeholders (section 6.1). This pattern addresses potential criticisms of services and products preemptively, while avoiding absolute claims or oversimplification. Through these two combinations, companies engage in strategic impression management by carefully crafting narratives around contentious sustainability issues, thereby mitigating potential stakeholder concerns while preserving operational autonomy. They collectively demonstrate a systematic approach to balancing transparency and operational flexibility.

Regarding engagement with secondary stakeholders, scholarship on corporate sustainability discourse has identified that Pronounce and Entertain (e.g., “we determined…could…”) signal companies’ forward-looking statements and determinations [9,21,23,25,27]. Building on this, we found that AI companies employed the combination of Strategic Pronounce and Strategic Entertain to announce long-term ethical roadmaps, aiming to project their societal commitments (section 6.4). Through strategically foregrounding definitive determination while hedging concrete outcomes, companies can minimise social accountability if their commitments fail to materialise. Similarly, previous research has shown that the use of “but” alongside negative adjectives can justify unsatisfactory CSR performance [22]. Instead, our analysis reveals how the co-occurrence of Strategic Counter (e.g., “but”) and Strategic Entertain promoted scientific strategies that informed the public about the potential to overcome assumed difficulties (section 6.5). This rhetorical pattern is critical in the statements, as the public increasingly expects transparency about how companies manage potential risks associated with the “black box” nature of AI technologies. In response, this targeted engagement pattern serves to position companies as engaged with both technical and public dimensions of AI development.

In addition, AI ethics statements also adopted distinctive engagement patterns to engage with secondary stakeholders. For example, Strategic Pronounce with Strategic Deny broke traditional one-sided narratives by offering balanced ethical perspectives (section 6.4). This rhetorical combination allows companies to present their vision as the definitive direction for global AI development. The combination of Strategic Counter and Strategic Deny (section 6.5), while common in professional communication to commonly build rapport [8,34,38], served a unique pragmatic function here. Companies use it to counter AI misunderstandings while building companies as authoritative voices on AI’s trajectory. Through these combinations, AI companies present themselves not merely as technology providers but as architects of AI’s global future, defining what AI should become and how society should understand it.

7.3. Practical implications

Corporate sustainability communication focuses on developing value propositions that address the needs of various stakeholders, particularly regarding social impact [84,85]. As ethical AI practices are central to this communication, reporting on responsible AI enhances corporate sustainability efforts. The above findings provide valuable insights for industry professionals who both produce and analyse such reports.

AI technology evolves rapidly and carries inherent risks. As a result, primary stakeholders, including clients handling sensitive data, third-party developers building systems, and regulators overseeing compliance, demand concrete safeguarding strategies before adopting these technologies [4,16]. Leading AI companies have adopted a hedged and evasive communication approach through targeted engagement strategies (e.g., Operational Counter and Operational Deny). Rather than promoting speculative breakthroughs or distant possibilities, they choose to emphasise proven achievements, existing service capabilities, and realistic near-term developments. This rhetorical approach serves multiple strategic purposes. It is designed to prevent customers and third-party developers from forming unrealistic expectations, while shielding companies from over-commitment risks and preserving discretion over future strategic directions. For the broader primary stakeholder group, these deliberate restraints function to suggest that these companies claim to prioritise accountability over hype. By grounding their communications in defined capabilities and limitations rather than ambitious promises, AI firms aim to build credibility with a diverse audience. They strategically provide regulators with evidence of compliance, while also indicating to investors the risks and opportunities that shape investment decisions.

Although secondary stakeholders are not direct participants in commercial transactions, they exert significant influence by scrutinising corporate ethical performance [5]. They demand commitments to transparency in AI decision-making processes and expect companies to prevent algorithmic bias that could perpetuate social inequalities [2,30]. Since AI technologies have a broad influence on society, companies must proactively address these concerns to avoid public backlash [13]. Leading AI companies have chosen to respond with a more ambitious yet calculated approach (cf. [9]). They communicate definitive long-term ethical standards and roadmaps, and they actively signal normative trajectories for AI technologies and their underlying principles through targeted engagement. These rhetorical tactics preemptively correct misconceptions and prevent misinformation from filling knowledge gaps. However, while companies assert these ethical commitments decisively, concrete outcomes and implementation details remain notably absent. By communicating proactively, nonetheless, companies represent themselves as industry leaders rather than defensive reactors, projecting social responsibility beyond mere compliance. This strategic positioning is designed to enable readers to see evidence of sophisticated risk management, operating on the premise that industry leaders who visibly address ethical concerns face fewer regulatory and reputational risks.

This tiered framework reveals how leading AI companies have shaped their AI ethics communication through strategic honesty. They disclose achievements alongside limitations using carefully structured engagement patterns. As Floridi [84] suggests, this approach is seen as a form of strategic stance management that enhances perceived strengths while downplaying weaknesses. Recognising this critical perspective, identifying these patterns nonetheless provides practical insights into this increasingly important genre of corporate sustainability discourse. Report writers can observe how established firms employ certain tones and selective transparency. Specifically, emerging companies can decide whether to adapt or reject these strategies when developing their own communication approaches. Readers, particularly investors, gain tools to better understand corporate rhetoric. When companies use cautious language, for instance, they can identify gaps between rhetoric and reality beneath polished narratives. Our analysis, therefore, positions strategic honesty not as an ethical ideal but as an observable, currently used business-driven communication approach. By understanding these rhetorical patterns, industry stakeholders can be encouraged to pursue more transparent communication practices.

7.4. Limitations and future work

This paper has two main limitations. Firstly, examining standalone corporate AI ethics statements may lead to overlooking relevant information presented in annual and ESG reports. More specifically, we only analysed a corpus of 122 statements, reflecting the nascent nature of AI ethics reporting. Larger corpora enable systematic identification of discourse patterns in professional discourse that smaller datasets may overlook [86]. As ESG reporting has become increasingly standardised, responsible AI performance reporting is likely to follow similar patterns [16]. This standardisation coincides with growing demand for AI ethics disclosures. Investigating engagement strategies thus becomes essential for understanding this emerging discourse genre.

Future research should expand the corpus to include not only standalone AI ethics statements but also relevant sections from other corporate reports. As AI ethics commitments are increasingly integrated into broader corporate sustainability disclosures, examining these diverse sources would provide a more comprehensive dataset. This approach offers multiple benefits. From a theoretical perspective, it will contribute to discourse analysis in CSC by identifying and verifying engagement patterns across different reporting channels. In practice, collecting data from multiple sources will provide more evidence-based communication strategies for organisations seeking to improve stakeholder engagement in reporting their AI practices. The enhanced engagement strategies will foster a more sustainable governance dimension of CSR/ESG reporting. Both companies’ credibility and alignment with international standards will thus be enhanced. On the other hand, understanding these strategies will equip readers to evaluate corporate CSR/ESG practices more effectively. Investors gain particular value from interpreting AI ethics engagement, as it reveals risk management strengths and weaknesses that shape investment choices.

The expansion of corporate AI ethics statements will enable more robust corpus-based analyses of inter-genre differences. By employing both qualitative and quantitative approaches, researchers potentially identify that multiple genres differ systematically in how they engage primary and secondary stakeholders. These insights can inform the drafting of genre-appropriate disclosures and ultimately enhance the quality of stakeholder communication.

Secondly, given the significance of AI ethics disclosures in corporate reporting systems, more nuanced discourse analysis of engagement is necessary. Scholars could explore other rhetorical appeal strategies to understand how companies attempt to persuade or engage stakeholders. For example, analysing metadiscourse (ethos, pathos, and logos) [44] and multimodal engagement features could likewise provide a comprehensive understanding of how internationally influential AI companies attempt to retain their audiences. These discourse analyses would advance engagement research in pragmatics by identifying genre-specific features and providing the industry with concrete strategies to enhance transparency in sustainability communication.