Open Weights, Life Sciences And The Risk Of Enterprise AI Dependency

Harini Gopalakrishnan, Founder of theHaze.ai & Industry GTM, Vespa,ai.

gettyMore than 270 companies and organizations recently signed the “Open Weights and American AI Leadership” letter, including major model developers, technology companies and infrastructure providers. The argument there is that open-weight models can widen access, improve competition and give organizations greater control over how and where they deploy AI.

​However, Anthropic has remained a notable exception. Anthropic’s stance is that while open-weight models without dangerous capabilities are a public good, sufficiently capable models can create different risks. Once weights are released, safeguards can be removed and access cannot simply be withdrawn. Anthropic, therefore, is pushing the need for mandatory safety testing for sufficiently capable models, whether open or closed, a kind of nuanced approach to supporting blanket open models.

​While I agree with the need for safety testing, I might disagree with the motive behind that support—how does a model leader decide whether a model is safe or unsafe to release as open weights?

1. Who decides when a model is too dangerous to open?

2. How long should intelligence remain closed?

3. How do frontier model developers protect their investments in capabilities?

​Life sciences has dealt with this tension for decades: balancing investment in capability with medicine access for the larger public good and improvements. Drug discovery requires substantial investment, long development cycles and significant risk. Companies are therefore given a period of commercial exclusivity to reward innovation and recover their R&D investment. But that exclusivity is not permanent. Over time, patents expire and generic competition allows the benefits of scientific progress to spread more broadly.

I want to draw the same parallel with AI.

While not exactly identical, the underlying principle is the same: A model with dangerous capabilities does not suddenly become safe simply because a newer generation has been developed.

​Frontier AI companies should be able to monetize the latest generation of models. But as the frontier moves forward, previous generations should have a pathway toward broader availability, including open-weight release where appropriate, once they clear an independent safety assessment.

​I think of this as “regulated diffusion”: Restrict genuinely dangerous capabilities where necessary, but do not let temporary safety controls automatically become permanent control over accumulated intelligence.

Basically, reward innovation and validate risk independently.

But who is responsible for this safety assessment?​ ​That’s where I differ most with the current framing.

​Model developers clearly need to participate in safety evaluation. They understand their systems better than anyone else.

​But they cannot be the only ones making the determination.

​A company has an incentive to demonstrate that its model is safe enough to commercialize. It may also have an economic incentive to argue that a highly capable model is too risky to release openly. Both incentives can coexist with genuine safety concerns.

​That is why we need common benchmarks and independent assessment involving model developers, researchers, standards bodies, regulators and domain experts. This needs to be a consortium that looks at both open and closed models with the same rigor.

​A closed model is not automatically safe. An open model is not automatically dangerous.

A downloadable model creates different risks from an API-controlled model. But the underlying assessment of what the model can actually do should not depend on who owns it.

​There is another side to this debate that I think matters enormously to enterprise leaders.

​Satya Nadella has talked about organizations compounding human capital with what he calls “token capital,” with the goal of helping enterprises build their own AI capability rather than simply consuming AI. Microsoft describes that future as one where companies turn their knowledge, workflows and judgment into AI systems that continuously improve.

​Are enterprises accumulating AI capability, or accumulating AI dependency?

​Every major AI investment should leave something behind that the organization owns. That could be institutional knowledge, evaluations, retrieval systems, domain context, learning loops, business rules, workflows or intellectual property. If the model gets better next year, those assets should make the enterprise better too. If, instead, every new capability remains tied to one provider’s models and proprietary services, the enterprise may consume increasing amounts of intelligence without building much capability of its own.

​This is not an argument against closed frontier models. Enterprises should use them when they create differentiated value. Nor is it an argument that enterprises should operate every model themselves. It is simply about understanding the difference between renting intelligence and building capability.

​Therefore, as AI investments grow, I think executives should periodically ask three questions.

​1. What AI capability are we actually accumulating?​

After all this AI spending, what does the enterprise actually own that compounds?​ After the project is complete, what knowledge, evaluations, workflows, context or learning remains inside the organization?​

2. Could we move that capability to another model tomorrow?

If the underlying model changes, can the enterprise carry its accumulated knowledge and capability forward, or does much of it disappear with the vendor?

​3. Are we accepting a vendor’s definition of risk, or benchmarking it independently?

Is the enterprise evaluating performance through its own benchmarks and independent standards?

​The open-weight debate will continue because both sides have legitimate interests. But wider access and enterprise control matter too.

At the industry level, the goal should be regulated diffusion: Reward the innovator, independently validate risk, and allow proven capabilities to spread when it is responsible to do so.

​Use the best intelligence available today. But distinguish between the intelligence you consume and the capability you build.

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?