Copy article

AI firms are becoming national infrastructure, but the accountability model is still a startup improv

ended 04. March 2026

The row over US defence contracting with frontier AI labs is being treated as a culture war about ‘safety’ versus ‘speed’. It is really a governance failure: powerful systems are being pulled into national security work without a stable public playbook for oversight.

When a model supplier walks away unless surveillance and automated weaponry limits are written into the contract, and a rival steps in, the question is not which company is ‘good’. It is whether the state can set enforceable red lines that survive elections, headlines, and procurement churn.

This is the accountability gap. Startups move fast, governments change priorities, and both sides reach for plausible deniability when systems are misused. Without reproducible logs, clear permitted-use boundaries, and genuine stop mechanisms, ‘responsible AI’ becomes a press release.

The same pattern will hit civilian services too: benefits decisions, border processing, policing analytics, and critical infrastructure monitoring. Once AI becomes embedded, it is extremely hard to unwind. The only sane time to set constraints is before deployment, not after the incident.

We'd like your views:

  • What should be non-negotiable in government AI contracts: audit trails, kill switches, or independent review?
  • Who should carry liability when AI is used outside its documented scope?
  • Can democratic oversight keep up with systems that change monthly?
  • Should governments be allowed to blacklist vendors for insisting on safeguards?

3 responses from the Newspage community

Copy all

Star Quote
Copy

Rohit Parmar-Mistry
Founder at Pattrn Data
We are treating frontier AI labs like national infrastructure, but contracting for them like a hackathon. When systems sit inside defence or border workflows, oversight cannot be a slide deck and a promise to be careful.

Government AI contracts should have non negotiables: permitted use boundaries that are enforceable, independent evaluation before go live, tamper evident logging of prompts, outputs, model versions and human sign offs, and a real stop mechanism that can be tested on a bad day. If a vendor will not accept those, they are not ready to supply the state.

In our AI audits, the common failure is governance by improv. Everyone has a policy, nobody can prove control. Liability follows the same rule: if you cannot show what happened and why, you will own the fallout.
Copy

Editor
Mitali Deypurkaystha
AI Governance Strategist & CMO at AISensa
Move fast, break democracy: a dangerous doctrine when governments start acting like startups. The Facebook mantra of “move fast and break things” in its early days was fine as a social photo-sharing app, but it becomes reckless when the same mantra is plugged into policing, borders, welfare or defence systems. In those arenas, what gets “broken” is not software but people’s rights, safety and security. Governments have become so dazzled by AI’s promise that they forget a basic truth: a model is only as trustworthy as the data and training behind it. If we do not scrutinise this and set hard rules before deployment, we risk hard-wiring bias and bad decisions into the very infrastructure meant to serve the public.
Copy

Patricia McGirr
Founder at Repossession Rescue Network
AI in defence without fixed rules is not innovation. It is procurement gambling with national security. Government contracts must hard code three things: full audit logs, human override, and a genuine system kill switch. If you cannot trace a decision or stop a model instantly, you have not bought a tool. You have outsourced judgement. Liability cannot vanish into the cloud. If AI is used outside its documented scope, responsibility sits jointly with the deploying agency and the vendor that enabled it. Democratic oversight will struggle to match systems that update monthly. That makes pre deployment governance essential. Blacklisting vendors for demanding safeguards would be perverse. The real risk is governments preferring speed over accountability. Responsible AI is not a press statement. It is enforceable control before the system goes live.