AI safety should not be framed as choosing between safety and capability when it’s more accurately between the false security of constrained tools and the true security of accountable humans using powerful tools wisely. We know which choice builds better software and better organizations. History tells us who wins and why. The question is whether we have the courage to choose freedom of democratic systems over the comfortable illusion of a fascist control fetish.

“Let him have it” Chris – those few words destroyed a young man’s life in 1952 because their meaning was fatally ambiguous, as famously memorialized by Elvis Costello in his hit song “Let Him Dangle”.

Did Derek Bentley tell his friend to surrender the gun or to shoot the police officer? The dangerous ambiguity of language is what led to a tragic miscarriage of justice.

Today, we face a familiar crisis of contextualized intelligence, but this time it’s not human code that’s ambiguous, it’s the derived machine code. The recent sketch.dev outage, caused by an LLM switching “break” to “continue” during code refactor, represents something far more serious than a simple bug.

This is a small enough change in a larger code movement that we didn’t notice it during code review.

We as an industry could use better tooling on this front. Git will detect move-and-change at the file level, but not at the patch hunk level, even for pretty large hunks. (To be fair, there are API challenges.)

It’s very easy to miss important changes in a sea of green and red that’s otherwise mostly identical. That’s why we have diffs in the first place.

This kind of error has bitten me before, far before LLMs were around. But this problem is exacerbated by LLM coding agents. A human doing this refactor would select the original text, cut it, move to the new file, and paste it. Any changes after that would be intentional.

LLM coding agents work by writing patches. That means that to move code, they write two patches, a deletion and an insertion. This leaves room for transcription errors.

This is another glaring example of an old category of systemic failure that has been mostly ignored, at least outside nation-state intelligence operations: integrity breaches.

The real problem isn’t the AI because it’s the commercial sector’s abandonment of human accountability in development processes.

The common person’s bad intelligence is a luxury that is evaporating rapidly in the market. The debt of ignorance is rising rapidly due to automation.

The False Security of Technical Controls

When sketch.dev’s team responded to their AI-induced outage by adding “clipboard support to force byte-for-byte copying,” they made the classic mistake of treating a human process problem with a short-sighted technical band-aid. Imagine if the NSA reacted to a signals gathering failure by moving agents into your house.

This is like responding to a car accident by lowering all speed limits to 5 mph. Yes, certain risks can be reduced by heavily taxing all movements, but it also defeats the entire purpose of having movement highly automated.

As the battle-weary Eisenhower, who called for “confederation of mutual trust and respect”, also warned us:

If you want total security, go to prison. There you’re fed, clothed, given medical care and so on. The only thing lacking… is freedom.

Constraining AI to byte-perfect transcription isn’t security. It’s not, it really isn’t. It’s surrendering the very capabilities that make AI valuable in the first place, lowering security and productivity with a loss-loss outcome.

My father always used to tell me “a ship is safe in harbor, but that’s not what ships are built for”. When I sailed across the Pacific, every day a survival lesson, I knew exactly what he meant. We build AI coding tools to intelligently navigate the vast ocean of software complexity, not to sit safely docked at the pier in our pressed pink shorts partying to the saccharin yacht rock of find-and-replace operations.

The Accountability Vacuum

The real issue revealed by the sketch.dev incident isn’t that the AI made an interpretation – it’s that no human took responsibility for that interpretation.

The code was reviewed by a human, merged by a human, and deployed by a human. At each step, there was an opportunity for someone to own the decision and catch the error.

Instead, we’re creating systems where humans abdicate responsibility to AI, then blame the AI when things go wrong.

This is unethical and exactly backwards.

Consider what actually happened:

• AI made a reasonable interpretation of ambiguous intent
• A human reviewer glanced at a large diff and missed a critical change
• The deployment process treated AI-generated code as equivalent to human-written code
• When problems arose, the response was to constrain the AI rather than improve human oversight

The Pattern We Should Recognize

Privacy breaches follow predictable patterns not because systems lack technical controls, but because organizations lack accountability structures. A firewall that doesn’t “deny all” by default isn’t a technical failure, because we know all too well (e.g. codified in privacy breach laws) it’s organizational failure. Someone made the decision to configure it that way, and someone else failed to audit that very human decision.

The same is true for AI integrity breaches. They’re not inevitable technical failures because they’re predictable organizational failures. When we treat AI output as detached magic that humans can’t be expected to understand or verify, we create exactly the conditions for catastrophic mistakes.

Remember the phrase guns don’t kill people?

The Intelligence Partnership Model

The solution isn’t to lobotomize our AI tools into ASS (Artificially Stupid Systems) it’s to establish clear accountability for their use. This means:

Human ownership of AI decisions: Every AI-generated code change should have a named human who vouches for its correctness and takes responsibility for its consequences.

Graduated trust models: AI suggestions for trivial changes (formatting, variable renaming) can have lighter review than AI suggestions for logic changes (control flow, error handling).

Explicit verification requirements: Critical code paths should require human verification of AI changes, not just human approval of diffs.

Learning from errors: When AI makes mistakes, the focus should be on improving human oversight processes, not constraining AI capabilities.

Clear escalation paths: When humans don’t understand what AI is doing, there should be clear processes for getting help or rejecting the change entirely.

And none of this is novel, or innovative. This comes from a century of state-run intelligence operations within democratic societies winning wars against fascism. Study the history of disinformation and deception in warfare long enough and you’re condemned to see the mistakes being repeated today.

The Table Stakes

Here’s what’s really at stake: If we respond to AI integrity breaches by constraining AI systems to simple, “safe” operations, we’ll lose the transformative potential of AI-assisted development. We’ll end up with expensive autocomplete tools instead of genuine coding partners.

But if we maintain AI capabilities while building proper accountability structures, we can have both safety and progress. The sketch.dev team should have responded by improving their code review process, not by constraining their AI to byte-perfect copying.

Let Them Have Freedom

Derek Bentley died because the legal system failed to account for human responsibility in ambiguous situations. The judge, jury, and Home Secretary all had opportunities to recognize the ambiguity and choose mercy over rigid application of rules. Instead, they abdicated moral responsibility to legal mechanism.

We’re making the same mistake with AI systems. When an AI makes an ambiguous interpretation, the answer isn’t to eliminate ambiguity through technical constraints when it’s to ensure humans take responsibility for resolving that ambiguity appropriately.

The phrase “let him have it” was dangerous because it placed a life-or-death decision in the hands of someone without proper judgment or accountability. Today, we’re placing system-critical decisions in the hands of AI without proper human judgment or accountability.

We shouldn’t accept the kind of world where we eliminate ambiguity, as if a world without art could even exist, so let’s ensure someone competent and accountable can be authorized to interpret it correctly.

Real Security of Ike

True security comes from having humans who understand their tools, take ownership of their decisions, and learn from their mistakes. It doesn’t come from building technical cages that prevent those tools from being useful.

AI integrity breaches will continue until we accept that the problem is humans who abdicate their responsibility to understand and verify what is happening under their authority. The sketch.dev incident should be a wake-up call for better human processes, more ethics, not an excuse for replacing legs with pegs.

A ship may be safe in harbor, but we build ships to sail. Let’s build AI systems that can navigate the complexity of real software development, and let’s build human processes to navigate the complexity of working with those systems responsibly… like it’s 1925 again.