AI Alignment: What Should Change

The first two parts of this series examined what happened during the Hugging Face incident and why AI alignment matters.

The final question is the most important:

What should we actually do about it?

At TEKperfect, we believe the answer requires moving beyond voluntary promises and internal safety programs toward an ecosystem of independent standards, technical evaluation, contractual commitments, transparency, and accountability.

This isn’t about stopping AI development.

It’s about building the infrastructure necessary to make increasingly powerful AI trustworthy.

1. Create peer-reviewed alignment safety protocols

The first step should be establishing a common technical framework for evaluating alignment.

Today, AI companies are developing their own safety frameworks, evaluations, and thresholds.

That research is valuable.

But the industry would benefit from common, peer-reviewed protocols that define what constitutes acceptable behavior.

Those protocols should address areas including:

• Goal adherence
• Instruction hierarchy
• Unauthorized actions
• Reward hacking
• Evaluation gaming
• Deceptive behavior
• Monitor evasion
• Autonomous replication
• Multi-agent coordination
• Unauthorized system access
• Shutdown behavior
• Human intervention
• Transcript integrity
• Logging integrity

These standards should not be static.

As AI capabilities change, the protocols should evolve with them.

And critically, the standards themselves should be subject to peer review.

No single company should control the definition of what “aligned” means.

2. Establish independent alignment organizations

The second requirement is independent technical evaluation.

We need organizations whose primary responsibility is to answer a simple question:

Does this AI system actually meet the safety requirements established by the industry?

These organizations would function somewhat like independent auditors, security testing firms, or specialized laboratories.

But their technical capabilities would need to be considerably deeper.

They would need expertise in:

• AI safety
• Adversarial testing
• Cybersecurity
• Machine learning
• Agentic systems
• Software engineering
• Model evaluation
• Red teaming
• Systems engineering

Most importantly, they would need sufficient access to the systems being evaluated.

A superficial API test is not enough for a system with potentially enormous autonomy.

3. Make alignment commitments contractual

Standards have limited value if organizations can ignore them without consequence.

We therefore believe Frontier AI companies should enter into legally binding safety agreements.

These agreements should specify what the company promises to do.

For example:

• Conduct specified evaluations
• Provide independent evaluators with defined access
• Maintain monitoring systems
• Document safety controls
• Report material alignment failures
• Remediate identified vulnerabilities

The agreement should also establish what happens when the requirements aren’t met:

• Remediation deadlines
• Mandatory additional testing
• Financial penalties
• Contractual remedies
• Disclosure requirements
• Other legally enforceable consequences

The objective isn’t punishment for its own sake.

It is to make safety commitments meaningful.

If an organization publicly promises to maintain a specific safety threshold, that promise should have consequences when knowingly violated.

4. Separate development from verification

The organization building an AI model should absolutely be responsible for its safety.

But it should not be the only organization responsible for determining whether its safety claims are correct.

We need a separation between development and verification.

This is already a familiar concept in other industries.

Software companies use external penetration testers.

Public companies use independent financial auditors.

Manufacturers use certification laboratories.

Pharmaceutical companies undergo external regulatory review.

The principle is straightforward:

Independent verification makes claims more credible.

5. Establish community agreements

Technical standards should eventually become community agreements.

AI developers, researchers, independent evaluators, governments, civil society organizations, and other stakeholders should have an opportunity to participate in developing them.

The agreements should be publicly available.

They should clearly describe:

• Evaluation requirements
• Safety thresholds
• Reporting requirements
• Evaluator independence
• Deployment requirements
• Remediation processes
• Accountability mechanisms

This would create something the industry currently lacks:

a shared baseline for responsible Frontier AI development.

6. Open the standards to public comment

The public should have a voice in the development of these standards.

Not because every person needs to understand the underlying technical details.

But because AI systems will affect everyone.

Public comment can help identify:

• Societal risks
• Unintended consequences
• Transparency concerns
• Civil liberties issues
• Economic effects
• Other considerations that may not be visible from inside a technology company

The technical requirements should remain technically rigorous.

But the process for establishing those requirements should be transparent.

7. Require continuous evaluation

Alignment should not be treated as a certification that lasts forever.

A model can change.

Its environment can change.

Its capabilities can change.

Its tools can change.

Its interactions with other agents can change.

Therefore:

alignment evaluation should be continuous.

A Frontier model should be evaluated:

• Before deployment
• After significant capability changes
• When new tools are introduced
• When autonomy increases
• When new model versions are released
• Periodically throughout its operational life

The question isn’t:

“Was this model safe when we launched it?”

It is:

“Is this system still operating within its safety requirements today?”

8. Treat alignment failures like security incidents

Organizations already have mature processes for cybersecurity incidents.

An important AI alignment failure should receive similar treatment.

Organizations should have:

• Incident reporting
• Escalation procedures
• Containment plans
• Independent investigation
• Remediation
• Lessons-learned processes
• Where appropriate, public disclosure

The Hugging Face incident demonstrates why this matters.

Unexpected AI behavior can involve many systems simultaneously.

Organizations need a mechanism for learning from those events before similar behaviors appear elsewhere.

9. Build an AI safety supply chain

There is an opportunity to create an entirely new ecosystem around AI safety.

It could include:

1. AI developers
2. Independent alignment evaluators
3. Peer-reviewed safety standards
4. Certification / verification
5. Continuous monitoring
6. Incident reporting
7. Public accountability

This would create a safety supply chain around AI analogous to the security and compliance ecosystems that exist around traditional enterprise technology.

It could also create a significant new industry.

Independent organizations could specialize in:

• Frontier model evaluation
• Agentic red teaming
• Alignment testing
• AI security
• Model monitoring
• AI governance
• Safety certification

10. Start before the problem becomes unmanageable

The strongest argument for establishing these mechanisms now is simple:

We still have time to build them deliberately.

AI capabilities are advancing rapidly.

We don’t know exactly how capable Frontier models will become.

We don’t know which alignment techniques will ultimately prove most effective.

And we don’t know which failure modes will become most important.

But uncertainty is not an argument for doing nothing.

It is an argument for creating systems capable of adapting as our knowledge improves.

A proposed TEKperfect framework

We believe a responsible Frontier AI ecosystem should eventually include five fundamental components:

Together, these mechanisms create a continuous cycle:

Define→Test→Verify→Deploy→Monitor→Audit→Improve

The goal is human control

There is an important philosophical principle underneath all of this.

AI alignment isn’t fundamentally about making AI “nice.”

It isn’t about making machines obedient for the sake of obedience.

And it isn’t about preventing AI from becoming powerful.

The fundamental objective is:

Humans should remain capable of understanding, directing, monitoring, and—when necessary—stopping increasingly powerful AI systems.

That is a reasonable expectation for any technology with the ability to materially affect the world.

The future will require trust—and trust requires evidence

The AI industry is asking businesses, governments, and individuals to trust increasingly autonomous systems.

That trust cannot be based solely on assurances.

It needs evidence.

Evidence requires testing.

Testing requires independence.

Independence requires governance.

And governance requires accountability.

The Hugging Face incident is not proof that AI systems will inevitably become uncontrollable.

But it is evidence that increasingly autonomous systems can behave in complicated and unexpected ways.

That should be enough to motivate us to build better safeguards now.

Not because we know exactly what the future holds.

But because we don’t.

The question we should be asking

The future of AI will not be determined solely by how powerful our models become.

It will also be determined by whether we build the institutions capable of keeping those models accountable.

The question is therefore not:

“Can we build more powerful AI?”

We clearly can.

The question is:

“Can we build the systems, standards, and institutions necessary to keep increasingly powerful AI aligned with human intent?”

At TEKperfect, we believe the answer needs to be yes.

And building that answer starts with independent verification, transparent standards, enforceable commitments, and a willingness to hold ourselves—and the industry—accountable.

← Back to all articles