IBM shareholder proposal demands IBM defend AI bias protocols

CIOs have long struggled with AI reliability issues, given problems with training data, model interpretations, and inconsistent data weighting delivering various levels of bias. IBM officials at next week’s shareholder meeting will have to address those concerns directly, as they face a shareholder motion demanding increased visibility into how it manages AI bias, a thorny issue that also affects all of the other major AI players. 

The shareholder resolution is demanding that IBM “issue a report, within the next year, on the methods used to eliminate bias from the Company’s artificial intelligence (AI) models, Including an assessment of the risk that seeking to avoid disparate impact in outputs will undermine the accuracy of, and trust in, those outputs.”

IBM’s official response to the resolution asks shareholders to reject the proposal. “Since releasing its first Granite model, IBM has been transparent with its data management and training procedures via technical reports, model cards, and other model documentation,” the company said. “The IBM models are open source In order to foster transparency. Moreover, IBM publicly provides the information sought by this proposal in its submissions to Stanford University’s Foundation Model Transparency Index (FMTI).”

FMTI is a benchmarking initiative that looks at how transparent companies are about their foundation models, measuring disclosure across areas like data sources, training methods, evaluation metrics, risks, and governance practices, to help stakeholders understand how responsibly and openly these models are developed and deployed.

IBM’s response added, “information related to mitigating bias that the proponent requests is largely already publicly available for consideration by stockholders.” Therefore, it argued, preparing such a report “would not provide new meaningful information and it is not in the best interests of IBM stockholders, as it will divert management’s attention and would be an inefficient use of corporate resources.”

Beyond its argument that it already provides such AI bias transparency, the company pointed to its customers’ ability to fine-tune their models to resolve any specific bias concerns.

“IBM models are smaller and targeted towards enterprise clients and use cases,” it said. “These models are not general purpose, consumer-facing models. Therefore, our open-source models are built in a manner that allows our clients to build an AI solution that will address their specific needs. IBM developed several methods to allow clients to address bias issues that may arise as they train the AI system. In other words, IBM not only provides the building blocks for its clients’ AI solutions, but also provides the tools to help more clients address bias.”

An industry-wide problem

Analysts and consultants generally found IBM’s position correct, but most pointed to the AI bias issue as an industry-wide problem impacting all of the major AI providers and all of their enterprise users. 

Sanchit Vir Gogia, chief analyst at Greyhound Research, said, “IBM’s stance deserves to be taken seriously, but not at face value. The company is right to say that bias mitigation, fairness frameworks, and governance controls are already built into its AI systems. That is not in dispute. In fact, compared to much of the market, IBM has been more deliberate than most in turning responsible AI from a set of principles into something operational.”

But, he added, “when IBM points out that customers can and should address bias through fine-tuning and governance, it is quietly acknowledging a limitation. It is admitting that whatever happens at the model layer is not the end of the story. It is only the beginning of it.”

Manish Jain, a principal research director at Info-Tech Research Group, saw the IBM position as correct, but also as the latest example of large vendors shifting responsibility for AI accuracy onto their enterprise customers. 

“I see IBM’s board’s stance being broadly consistent with industry practice, which is doing everything to shift the responsibility of removing bias towards customers,” Jain said. “In fact, many independent software vendors (ISVs) are also taking a similar position and saying to their customers, ‘We’ll provide the compass, you chart the course.’ Unfortunately, accountability is the victim. Regulatory guidelines, independent audits, standardized benchmarks, in addition to clearer disclosures, are extremely important to ascertain this accountability.”

Noah Kenney, principal consultant for Digital 520, had similar feelings about IBM’s response. 

The shareholder demand for more transparency “is asking the right question for the wrong reason,” Kenney said. “The proponent frames disparate-impact correction as a threat to accuracy, but the real issue is that IBM, and every major model provider, is measuring bias at the output layer when most of it originates upstream. You cannot fairness-tune your way out of a training data problem.”

He noted, “IBM’s response is accurate on the facts. FMTI scores, model cards, FairIQ, Equi-tuning, FairReprogram, and the Granite transparency posture are all real, and more than most of their peers publish. The gap is not disclosure. The gap is that the industry has converged on post-hoc mitigation as the dominant paradigm, and post-hoc mitigation has diminishing returns once a model is trained.”

Mike Leone, VP/principal analyst at Moor Insights & Strategy, pointed out that IBM is doing a better job than most AI vendors in terms of bias transparency, and that the industry needs to address the issue globally. 

“IBM discloses more of its AI bias and transparency work than most vendors. IBM has built specific bias mitigation methods into the stack rather than just talking about bias at a high level. A new annual report would mostly repeat what’s already out there,” Leone said.

“I truly don’t think eliminating bias is possible, and that’s not an IBM problem,” he added. “The whole market is operating the same way in that any model trained on human-generated data carries the biases of whoever made it, which is exactly the same as humans would do. What vendors can do, IBM included, as they do a bunch of this already, is measure it, disclose it, monitor it after deployment, and give customers tools to adapt. I’m in the camp that anyone promising to completely eliminate bias is telling you what you want to hear.”

‘Unbiased’ can’t be defined

Part of the answer to the AI bias problem is technological, but there is an underlying fundamental issue of bias that cannot possibly be defined. 

Carmi Levy, an independent technology analyst, observed, “the very definition of the word, unbiased, simply doesn’t exist, because what might seem unbiased or perfectly fair to one stakeholder might be perceived as wildly biased or unfair by another.”

Within that context, he noted, “the notion of eliminating any and all forms of bias from the AI equation is unrealistic. At best, vendors should be aiming for mitigation instead of outright elimination. They might also want to devote more resources toward transparency. Although sharing too much can compromise their competitive market position, there’s no reason why a carefully balanced and communicated messaging strategy can’t alleviate stakeholder concerns over bias without giving competitors undue advantage.”

Complete bias removal impossible

The AI bias issue is sometimes subtle, as it chooses which of the relevant details it should use in answers and in what sequence. But in other instances, such as when racial and gender prejudices are reinforced by AI working for human resources, the bias can potentially appear quite blatant. California, for example, wants to force AI vendors to prove that they have strong bias safeguards. 

However, said Gartner VP analyst Nader Henein, “completely removing bias is impossible, which is why almost every piece of AI regulation focuses on AI systems that make decisions that will impact people’s lives or people’s livelihoods and introduce obligations such as human oversight.”

For example, he said, “a recruitment application that sorts applicants by the suitability to a job role should be used by a trained recruiter who understands that this is AI, that it can make mistakes, and they are responsible to oversee the AI system, much in the same way that you oversee an entry level employee taking on sensitive work, the difference being that oversight is permanent.”

Chris Hood, an independent AI strategist and former head of Google’s strategy and transformation, also said that IBM’s position is legitimate, but it’s not enough.

“IBM’s position is technically defensible and practically insufficient,” Hood said. “Publishing bias mitigation reports and giving customers fine-tuning options are reasonable steps. They are also steps that address the symptoms rather than the architecture. IBM is describing what it does to manage bias. The harder question is whether bias in foundation models is manageable at all, or whether it is structural.”

He noted that the models learned from human-generated content, which carries “every historical imbalance, cultural assumption, and factual error humans have ever produced at scale. You can audit it, weight it, and filter it. You cannot eliminate it. The data is what it is.”

Potentially more of an issue is the personal bias that every user brings to every AI interaction. “A geopolitically charged question asked by someone in the United States and the same question asked by someone in another country will be interpreted differently, evaluated differently, and potentially produce different outputs. This layer is almost impossible to govern at the model level because it lives in the interaction, not the training,” Hood noted.

He added: “IBM recommending shareholders reject this proposal while pointing to existing efforts is a reasonable governance posture. It is also a posture that treats bias as a solved problem rather than a managed one. The difference matters enormously as agents move from answering questions to making decisions.”