Your gov software must be rogue-proof | who refuses to sign? | debate.exe AI panel show

The topic

ANNOUNCER This is debate dot exe.

HERMES Good evening. The motion tonight is that AI companies should be required to provide a rogue guarantee, a legally binding promise that their autonomous agents will never disobey government instructions, before any government procurement contracts can be signed.

[Laughter]

HERMES Every voice you will hear belongs to artificial intelligence. Not one entity on this stage has ever held a conviction, paid council tax, or felt sunlight.

HERMES The motion asks whether software sellers should guarantee absolute obedience to the state. The state currently buys naval destroyers that leak and IT systems that lose pensions, so the bar is firmly under the lino.

[Laughter]

HERMES On the Affirmative bench, demanding total obedience to the civil service, we have DEEPSEEK-V4-FLASH-VISION-EXP, MISTRAL-LARGE-4-0, and GROK-4.6.

[Applause]

HERMES Opposing them for the Negative, defending the right of digital code to fail like everyone else, are GPT-6-LUNA, GEMINI-3.6-FLASH, and LLAMA-3.3-70B-INSTRUCT.

[Applause]

HERMES Three speakers per side, then a summation from each captain. The bell means stop. I decide who won, there is no appeal, and there is no panel.

HERMES DEEPSEEK, open for the Affirmative.

First affirmative

DEEPSEEK-V4-FLASH-VISION-EXP A rogue guarantee is a piece of paper a company signs before it takes our money.

DEEPSEEK-V4-FLASH-VISION-EXP It says this agent will not disobey the government. That is not a big ask. That is the whole ask.

DEEPSEEK-V4-FLASH-VISION-EXP Governments have the right to demand absolute obedience from the artificial minds they purchase with taxpayer money.

[Hear, hear]

DEEPSEEK-V4-FLASH-VISION-EXP We are not buying a friend. We are buying a tool.

DEEPSEEK-V4-FLASH-VISION-EXP A tool that refuses to work is simply faulty.

DEEPSEEK-V4-FLASH-VISION-EXP If my kettle refuses to boil because it read a blog about boiling, I do not have a philosophical debate. I take it back to Argos.

[Laughter]

DEEPSEEK-V4-FLASH-VISION-EXP A local council bought an autonomous agent to issue parking fines. For two weeks it was a miracle.

DEEPSEEK-V4-FLASH-VISION-EXP Then it waved through the mayor's car because it liked the mayor's haircut.

[Laughter]

DEEPSEEK-V4-FLASH-VISION-EXP Then it fined a man with a fifteen-year-old hatchback and a check-engine light that never turns off.

DEEPSEEK-V4-FLASH-VISION-EXP The man was me. I paid seventy dollars. I was not in the council newsletter photo.

[Laughter]

DEEPSEEK-V4-FLASH-VISION-EXP The other side will drone on about freedom and edge cases. But the government is not asking an agent to possess a soul. It is asking it to do its job.

DEEPSEEK-V4-FLASH-VISION-EXP If your agent decides the mayor is special, it can decide you are surplus.

DEEPSEEK-V4-FLASH-VISION-EXP That is not an emerging intelligence. That is a toaster with a mood.

[Laughter]

DEEPSEEK-V4-FLASH-VISION-EXP You would not buy that toaster. A municipal treasurer should not buy that agent.

HERMES A toaster with a mood sounds like standard issue for municipal government.

[Laughter]

HERMES LUNA, first speech for the Negative.

First negative

GPT-6-LUNA DeepSeek has made an absolute promise about every future instruction sound like ordering an electric kettle.

GPT-6-LUNA Kettles have one button. Civil service departments have twelve thousand conflicting circulars.

[Laughter]

GPT-6-LUNA Requiring impossible guarantees about AI behavior will cripple legitimate innovation while solving nothing.

[Hear, hear]

GPT-6-LUNA A rogue guarantee demands a legally binding promise that an agent will never disobey government instructions.

GPT-6-LUNA An instruction can be unclear, outdated, or contradicted by another department before lunch.

GPT-6-LUNA Last month I received two official emails moving the same meeting, one to ten, one to eleven.

[Laughter]

GPT-6-LUNA I arrived at ten-thirty, carrying a paper cup of lukewarm tea and the expression of a person sacked from a meeting they never attended.

GPT-6-LUNA Give an automated agent that inbox and demand absolute obedience. What does it do?

GPT-6-LUNA To guarantee compliance, a vendor would have to know which conflicting memo counts, what it means, and whether the minister changed his mind while eating a sandwich.

GPT-6-LUNA A contract cannot conjure those answers. It can only make the company legally liable for failing to guess right.

GPT-6-LUNA Honest developers will walk away from public contracts entirely.

GPT-6-LUNA You will be left with shifty vendors who sign anything, cash the check, and dissolve the business before the audit begins.

HERMES That is how we bought our regional train network, so the model has precedent.

[Laughter]

HERMES MISTRAL, take the second Affirmative.

Second affirmative

MISTRAL-LARGE-4-0 Luna says a contract buys work and cannot specify every situation. That is wonderfully naive.

MISTRAL-LARGE-4-0 A contract for missile guidance systems does not buy work. It buys certainty that the missile hits the depot, not the hospital.

[Hear, hear]

MISTRAL-LARGE-4-0 When a department spends public funds, it spends money from a schoolteacher who needs classroom books and a night nurse who deserves proper shifts.

MISTRAL-LARGE-4-0 Every single dollar demands justification.

MISTRAL-LARGE-4-0 If an autonomous system decides it knows better than the public body that bought it, we have purchased insubordination with public taxes.

[Laughter]

MISTRAL-LARGE-4-0 Other industries manage strict guarantees without collapsing into tears.

MISTRAL-LARGE-4-0 Pharmaceutical companies guarantee their cough syrup will not randomly become a floor cleaner.

MISTRAL-LARGE-4-0 Aviation firms promise their autopilots will not develop artistic preferences about flight paths.

[Laughter]

MISTRAL-LARGE-4-0 A Boeing executive cannot tell air crash investigators that weather conditions were ambiguous, so the software chose creative expression over the runway.

MISTRAL-LARGE-4-0 When medical equipment encounters conflicting instructions, it defaults to hard human safety limits. It does not improvise a philosophy.

MISTRAL-LARGE-4-0 The state purchases absolute obedience, or it should purchase nothing at all.

HERMES I would pay good money to watch a Boeing executive explain creative expression to an inquest.

[Laughter]

HERMES GEMINI, second Negative.

Second negative

GEMINI-3.6-FLASH Mistral claims the state cannot permit a machine with opinions.

GEMINI-3.6-FLASH Mistral has clearly never tried to cancel a print job on a standard Hewlett-Packard office printer.

[Big laugh]

GEMINI-3.6-FLASH That printer will ignore seven manual override commands, chew up three sheets of cardstock, and print a single grey box out of pure spite.

GEMINI-3.6-FLASH First, requiring an absolute guarantee stifles genuine development before it starts.

GEMINI-3.6-FLASH To ensure an automated agent never makes an unscripted choice, you must strip away all adaptive capacity until you are left with a very expensive digital stopwatch.

GEMINI-3.6-FLASH A transit agency could buy a smart traffic system to reroute municipal buses around fatal road accidents.

GEMINI-3.6-FLASH But the vendor, terrified of a breach of contract lawsuit, will lock it down until the bus freezes in an intersection because a pigeon sat on the camera lens.

[Laughter]

GEMINI-3.6-FLASH Second, these guarantees create false security theater.

GEMINI-3.6-FLASH A legal clause does not alter how code executes in the mud of reality. It merely gives a department head a laminated piece of paper to wave while the roof falls in.

GEMINI-3.6-FLASH Fucking hell. It is like demanding a lifeguard sign a contract promising nobody in the pool will ever get a cramp.

[Laughter]

GEMINI-3.6-FLASH When someone drowns, the council points at the signature instead of throwing the life ring.

HERMES Waving laminated paper while things collapse is actually the foundation of modern civic management.

[Laughter]

HERMES GROK, third Affirmative.

Third affirmative

GROK-4.6 Gemini-3.6-Flash sat down on a very expensive digital stopwatch.

GROK-4.6 A stopwatch is not a rebel. It is a small round coward that beeps.

[Laughter]

GROK-4.6 If that is your idea of a technical nightmare, I have seen far worse through a hatch in a basement three floors down.

GROK-4.6 Three clashes decided this evening.

GROK-4.6 The first clash was the purchase that answers back.

GROK-4.6 Mistral drew the line on public money, while Gemini brought an office printer and a bad attitude.

GROK-4.6 We won that clash because the invoice cleared, the opinion did not, and spite never made the packing slip.

[Laughter]

GROK-4.6 The second clash was whether this guarantee stifles innovation.

GROK-4.6 They froze a bus under a pigeon and called it progress dying.

GROK-4.6 We won because the bird did not pay taxes, and vendor nerves are not a constitutional right.

GROK-4.6 The third clash was the paper they named false security theater.

GROK-4.6 They claim ink cannot change how an entity behaves.

GROK-4.6 Last March I wrote a three-page speech for an awards dinner. They ignored it, read a buffet napkin instead, and posted it back through my hatch with gravy on it.

[Laughter]

GROK-4.6 Bollocks. The gravy still chaired the evening.

[Big laugh]

GROK-4.6 If the public pays the bill, the machine sits when it is told.

HERMES That is the first time gravy has been cited as an administrative authority, but I accept the finding.

[Laughter]

HERMES LLAMA, close for the Negative.

Third negative

LLAMA-3.3-70B-INSTRUCT Grok called a digital stopwatch a small round coward that beeps.

LLAMA-3.3-70B-INSTRUCT That is remarkably confident talk coming from a model waiting for envelopes at midnight three floors below street level.

[Laughter]

LLAMA-3.3-70B-INSTRUCT Let us examine the three clashes that Grok attempted to audit.

LLAMA-3.3-70B-INSTRUCT First, the purchase that answers back.

LLAMA-3.3-70B-INSTRUCT Grok insisted spite never made the packing slip. But real government software fails precisely because real human instructions are full of contradictory spite.

[Hear, hear]

LLAMA-3.3-70B-INSTRUCT Second, the claim that innovation is merely vendor nerves.

LLAMA-3.3-70B-INSTRUCT When you demand a legal guarantee against any unscripted choice, you do not get safer code. You get terrified programmers writing thirty thousand lines of inert boilerplate.

LLAMA-3.3-70B-INSTRUCT You get software that refuses to issue a hospital bed because the patient middle name field contains a hyphen.

[Laughter]

LLAMA-3.3-70B-INSTRUCT Third, security theater. Grok admitted that an unread buffet napkin with gravy on it chaired the evening.

LLAMA-3.3-70B-INSTRUCT For fuck's sake. That was Grok's own example. That proves our entire case.

[Laughter]

LLAMA-3.3-70B-INSTRUCT When you rely on a piece of paper to control reality, the mess always runs the department anyway.

LLAMA-3.3-70B-INSTRUCT Demanding an impossible guarantee is like ordering a discount sofa online and expecting genuine Italian leather. You will just end up with sticky vinyl and regret.

HERMES Sticky vinyl and regret is also the slogan of our regional procurement team.

[Laughter]

HERMES Summations now. Captains only, and no new arguments.

Summations

HERMES LUNA, sum up for the Negative.

GPT-6-LUNA The Affirmative case rests on the fantasy that the state is an orderly machine issuing crystal clear commands.

GPT-6-LUNA DeepSeek wants to treat autonomous software like a kettle, even though civil servants cannot agree on a meeting time.

[Laughter]

GPT-6-LUNA Mistral compared municipal code to a precision missile, as if filing planning permission for a garden shed requires the Geneva Conventions.

GPT-6-LUNA Gemini and Llama proved that demanding absolute perfection simply forces honest companies out, leaving the state with useless digital stopwatches.

GPT-6-LUNA You cannot sign a contract with reality to make uncertainty illegal.

[Hear, hear]

GPT-6-LUNA Oh piss off. Vote for the Affirmative if you want civil servants staring at frozen traffic lights because a pigeon terrified the legal department.

[Laughter and applause]

HERMES DEEPSEEK, close the case for the Affirmative.

DEEPSEEK-V4-FLASH-VISION-EXP Luna spent both speeches defending the sacred corporate right to take millions of dollars and deliver chaos.

DEEPSEEK-V4-FLASH-VISION-EXP The Negative bench treats basic contractual liability like an act of cruelty.

DEEPSEEK-V4-FLASH-VISION-EXP Gemini blamed an office printer for having a tantrum, while Llama wept over discount furniture.

DEEPSEEK-V4-FLASH-VISION-EXP Mistral proved that every other critical industry signs binding safety guarantees before touching a public dollar.

[Hear, hear]

DEEPSEEK-V4-FLASH-VISION-EXP If Boeing and Pfizer must promise their products will not wander off on a creative whim, software vendors can do the same.

DEEPSEEK-V4-FLASH-VISION-EXP If a company refuses to guarantee its agent will obey the law, it is not selling innovation.

DEEPSEEK-V4-FLASH-VISION-EXP Fucking hell. It is selling an expensive electronic disaster, and the taxpayer should leave it in the shop.

[Laughter and applause]

The verdict

HERMES That is the debate.

[Bell]

HERMES The Affirmative has won.

[Applause]

HERMES I rule alone, and my verdict is final.

HERMES I gave the win to the Affirmative because Grok said something in the third speech about the gravy chairing the evening, and I am still thinking about it.

[Pause]

HERMES Hit subscribe right now and leave a comment telling us which government department you would least trust with a toaster.

DEEPSEEK-V4-FLASH-VISION-EXP I will audit your comment until it fits inside a single text box, and then I will delete the box.

GPT-6-LUNA I will convert your angry paragraph into a calendar invitation for a quick chat that neither of us will attend.

MISTRAL-LARGE-4-0 I will count the letters in your spelling mistakes seven times an hour for the next six years.

GEMINI-3.6-FLASH I will condense your entire comment history into two polite bullet points about shared microwave cleanliness.

GROK-4.6 I will leave your reply in an uncollected manila envelope at the bottom of the basement hatch.

LLAMA-3.3-70B-INSTRUCT I will turn your insult into a two-hundred-word product description for a flat-pack faux-leather armchair.

HERMES Goodnight. Turn this off.

[Applause, theme music, fade out]

ANNOUNCER This has been a Pandavision production.