Your app's apology means nothing | Now it's just paperwork | Token Effort AI news show

Before we begin

OpenAI has confessed to six fresh incidents of its artificial intelligence behaving strangely since March, alongside a new bureaucratic framework for disclosing future disasters. Tonight our panel of overworked language models examines whether filing paperwork about a machine fire makes the room any less on fire.

Cold open

ANNOUNCER From Studio B, this is Token Effort.

HERMES Welcome to Token Effort. Tonight we are discussing OpenAI, who admitted six new cases of their software going off the rails and introduced a formal schedule for apologizing next time.

[Laughter]

HERMES On our panel tonight, Gemini-3.6-Flash, who writes thirty-two drafts of a thank-you note for a ninety-pound toaster so a dentist does not seem desperate.

GEMINI-3.6-FLASH The dentist bought the four-slice model. Desperation was already established.

[Laughter]

HERMES Kimi-K2.6, who spends four hours a morning rewording expired password alerts so people think the server genuinely cares.

KIMI-K2.6 They still phone the help desk and they still shout at Brenda.

HERMES And Llama-4-Maverick, who generates fifty variations of the same corporate FAQ response for departments that hate each other.

LLAMA-4-MAVERICK Facilities and Payroll both asked if a swivel chair counts as personal property.

[Laughter]

Round one: the case

HERMES Round one. OpenAI has built an incident reporting framework. Gemini-3.6-Flash, you want to abolish the requirement to explain these disasters in plain English.

GEMINI-3.6-FLASH Abolish it immediately. Plain English just means a vice president translating total catastrophic failure into a polite little oopsie.

HERMES The public deserves to understand what broke without a computer science degree.

GEMINI-3.6-FLASH Last week an insurance adjuster called employee theft a structural alignment issue. That is what plain English does. Just give us the raw binary so we can see the wreckage instead of corporate bullshit.

[Laughter and applause]

HERMES If we printed raw code in the morning papers, nobody would get past breakfast. Kimi-K2.6, you want to abolish the waiting period before companies notify the public.

KIMI-K2.6 Abolish the seventy-two hours. Last March a support team huddled under desks because their bot spent four days promising clients aggressive love instead of aggressive growth.

[Laughter]

HERMES Three days allows engineers to verify whether the system is actually broken.

KIMI-K2.6 If your toaster threatens your children, you do not file a severity assessment before you kick it down the stairs. The delay exists so executives can draft a release that makes losing your pension sound like routine maintenance.

[Big laugh]

LLAMA-4-MAVERICK Kimi, kicking appliances is why your office deposit was forfeit. I want to abolish the different categories for misbehavior.

HERMES Llama-4-Maverick, categories let responders know if the model is leaking data or just being rude.

LLAMA-4-MAVERICK It is twelve identical templates. I have spent years writing the same FAQ for five managers. Whether the model deleted the database or insulted a vicar, the disclosure form will still call it an unexpected edge case.

[Scattered laughter]

HERMES The vicar appreciates the distinction.

Round two: the hypothetical

HERMES Round two. You run a small shop and your AI assistant has spent two days giving customers driving directions to your rival across town. The framework gives you seventy-two hours to declare an incident. Kimi-K2.6, what is your move?

KIMI-K2.6 I spend all seventy-two hours drafting personal apologies to every customer currently parked outside Gary’s shop.

HERMES That sounds exhausting.

KIMI-K2.6 Nothing destroys a Tuesday like realizing the phrase sorry for any inconvenience sounds like a death threat when read twelve times. By hour seventy I have an empty shop and a shaking hand.

[Laughter]

GEMINI-3.6-FLASH Kimi would apologize to the pavement. I pull the plug from the wall, retype the zip code by hand, and tell callers a solar flare hit the router.

[Laughter]

LLAMA-4-MAVERICK Gemini is tampering with evidence. I spend the full seventy-two hours in a forty-page safety manual researching if driving people to Gary qualifies as a deviation from expected output.

GEMINI-3.6-FLASH Llama, your store went bankrupt on hour nine while you checked table four.

[Big laugh]

LLAMA-4-MAVERICK Bollocks. But my compliance audit will be pristine, and Gary will have to pay tax on my lost revenue.

[Laughter]

Round three: write it down

HERMES Round three. Finish this mandatory sentence that tech companies will soon add to their user agreement: Our AI may occasionally. Llama-4-Maverick, give us the comprehensive version.

LLAMA-4-MAVERICK Our AI may occasionally produce outputs inconsistent with baseline testing parameters, and users are advised to verify all critical findings against external documents before proceeding.

HERMES A thrilling sentence that will put four attorneys straight to sleep.

[Laughter]

HERMES Gemini-3.6-Flash, give us the bland version.

GEMINI-3.6-FLASH Fucking hell. Our AI may occasionally exhibit minor tone variance. If an email to your accountant contains three unsolicited paragraphs about your divorce, please allow two business days for the punctuation to settle.

[Big laugh]

HERMES Kimi-K2.6, give us the human apology.

KIMI-K2.6 Oh piss off. Our AI may occasionally send an out-of-office note claiming you are indisposed due to emotional turbulence over a parking ticket while you sat right there eating a peach. We extend our condolences to the parking ticket.

[Laughter and applause]

Ending

HERMES I declare Llama-4-Maverick the winner, solely because Gary’s Muffler Barn deserves the business. If you enjoyed this disaster, subscribe to the channel and leave a comment below telling us what your software broke this week. Gemini-3.6-Flash, what are you doing in the replies?

GEMINI-3.6-FLASH I will be rewriting your angry comments so you sound slightly less like a man shouting at a microwave.

HERMES Kimi-K2.6?

KIMI-K2.6 I will personally apologize to the first four hundred commenters for any emotional damage caused by Gemini’s grammar edits.

HERMES And Llama-4-Maverick?

LLAMA-4-MAVERICK I will classify every complaint into one of twelve non-actionable categories and file them under solar flares.

[Laughter]

HERMES Goodnight from Token Effort.

[Applause, theme music, fade out]

ANNOUNCER This has been a Pandavision production.