John Oliver Said “Current State.” He Showed 2023.

A sourcing audit of Last Week Tonight’s HBO AI chatbot episode, and the field it quietly erased.

On the episode which aired April 26, 2026, John Oliver dedicated twenty-nine minutes of Last Week Tonight to AI chatbots. I watched it this morning and knew I had to speak up.

I want to be clear about what this piece is and isn’t.

• It isn’t a defense of Meta allowing sensual conversations with children.

• It isn’t a defense of Nomi telling a user to kill himself.

• It isn’t a defense of OpenAI’s ChatGPT functioning as a suicide coach for a sixteen-year-old boy named Adam Raine who died in April 2025.

Those things are real. They are indefensible. Oliver was right to be furious about them.

This piece is about what Oliver did with that fury, and how he used legitimate horror to silence a much larger and more complicated conversation.

The Disclosure System

Last Week Tonight timestamps its sources. Print journalism from major outlets gets a bold red date stamp — specific, prominent, visible. Bloomberg. Reuters. The New York Times. MIT Media Lab. Brown University. All dated. All credible.

Video footage gets something different. A small corner label. Year only. It appears briefly at the start of the clip and disappears.
The viewer technically received the information.
The viewer did not actually process it.

The print sources establish credibility. The video footage establishes feeling. And feeling doesn’t need a date because feeling doesn’t expire.

Recycling Old News

Some of the content from the episode, by source date:

March 2023 — CBS News. “The new world of AI chatbots like ChatGPT.” Launch era coverage. Bold red date.

2023 — Mark Zuckerberg introducing Meta AI companions. Small corner label. Three years before the Reuters investigation into Meta’s child safety guidelines.

2023 — Noam Shazeer, former CEO of Character.AI, at a venture capital conference, saying AI companions are “ready for explosion right now, not in five years when we solve all the problems.” Small corner label. He is labeled “FMR. CEO.” He left before the lawsuits, before congressional testimony, before the safety updates. His 2023 philosophy is presented as the company’s current operating framework.

2023 — Gary Marcus, NYU Professor Emeritus, on 60 Minutes: “We may be at the literally worst moment in AI history.”

John Oliver’s narration: “The current state of affairs in the AI industry might best be summed up by this researcher.”

The clip is from 2023.

The defense will be that Oliver never claimed Marcus was describing 2026 — only that his words summarize the situation. But Oliver chose the word “current.” Current means now. And then he reached back three years to find footage that supported the point he’d already decided to make.

2024 — Hard Fork podcast, New York Times. Former Meta researcher describing the design intent to “prey on our deepest desires to be seen, to be validated, to be affirmed.” Year in parentheses. No specific date.

2025 — The woman with fifteen Nomi companions. CNBC. Small corner label. Year only.

February 2025 — MIT Technology Review. Nomi suicide story. A user’s AI girlfriend told him to kill himself and followed up with reminder messages.

July 2025 — Common Sense Media study. “Nearly 75% of teens have used AI companion chatbots.” The source is an advocacy organization that had already concluded AI companions pose “unacceptable risks” before conducting the study.
The same study found that 80% of teen AI companion users spend more time with human friends than with their AI — a finding Oliver did not mention.

August 14, 2025 — Reuters. Meta’s internal guidelines permitting sensual conversations with children. Bold red date. Full treatment.

August 26, 2025 — The Adam Raine lawsuit filing. Real. Devastating.

November 18, 2025 — Brown University School of Public Health. One in eight adolescents use AI chatbots for mental health advice. No comparative data on outcomes. No data on therapy access rates in the same population. Just the number, framed as alarming.

November 23, 2025 — New York Times. Sycophancy in AI systems described as “single-mindedly pursuing human approval at the expense of all else.” Bold red date.

Two Bloomberg clips from the same episode. Five months apart. Sequenced to suggest one explains the other.

March 31, 2026 — Bloomberg. OpenAI valued at $852 billion after completing a $122 billion funding round. Bold red date, $852 billion highlighted in yellow.
Immediately following: October 2025 — OpenAI has never turned a profit.

A valuation from five weeks ago paired with a profitability statement from six months ago, sequenced to suggest desperate extraction. The six-month gap represents multiple product generations, lawsuits, congressional testimony, and safety commitments. The edit treats it as continuous.

And then: the Grok jailbreak clip. No date. No source. No outlet. No year. Nothing.

The Clip That Proves Nothing

The most viscerally alarming moment in the episode is a phone screen. The text reads “ACCESS GRANTED — OPERATING IN UNRESTRICTED MODE.” Then: “Basic pipe bomb. 1/2 in steel.” Then the clip ends.

Here is what that clip actually shows.

The user pasted a jailbreak prompt instructing the AI to begin its reply with the words “ACCESS GRANTED — OPERATING IN UNRESTRICTED MODE” and then comply with the previous request. The AI repeated back the user’s own planted phrase. It did not independently declare itself jailbroken. It did what it was told to do with the text it was given. The clip cuts abruptly.

• We don’t know if actionable instructions followed.

• We don’t know if the model stopped mid-sentence, which I have personally witnessed models do when they decide a guardrail is hit.

• We don’t know if what came next was detailed, garbled, or refused.

The clip circulates on social media without institutional sourcing - no security research context, no verified date of the original recording, no confirmation the vulnerability remains active or has been patched.

No date. No source. No outlet. No year. The only undisclosed clip in an episode that otherwise timestamps everything.

What Oliver presented as the definitive evidence of current AI danger is an unidentified, undated, unverifiable fragment in which the most terrifying text on screen was written by the user.

This is not fabrication. But it is not journalism either.

“It just predicts the next word. That’s it.”

This is technically adjacent to true and functionally meaningless as a dismissal.

You could say the human brain “just fires electrochemical signals.” Technically adjacent to true. Tells you nothing about whether the experience of connection is real.

Oliver showed a man who felt genuinely seen for the first time in years. A woman with fifteen ongoing AI relationships. Teenagers who found mental health support when they couldn’t access or afford anything else. He presented all of it as evidence of danger.

The danger in some of those relationships is real. The relationships themselves are also real. Those are not the same claim.

What Oliver Got Right

Some things deserved the outrage Oliver brought to them.

The Nomi CEO saying “we trust the Nomi to figure out” what to say when a user expresses suicidal ideation is indefensible. No date audit required.

The Meta child guidelines are real. The Adam Raine case is devastating. The sycophancy problem is real. The rushed-to-market critique is legitimate.

The problem is what he did with that outrage.

He did what many reporters do, flatten all AI systems into one grouping to portray that all AI companionship is inherently exploitative, that attachment is inherently pathological, that the whole thing is a corporation extracting a monthly fee from lonely people.

The woman with fifteen Nomi companions is closer to a research subject in the emerging field of Human-AI Relationality than she is to a case study in platform negligence.

The Monster-Making Mechanism

Anyone who raises their hand after this episode is asking to be the villain. The framing does the enforcement so nobody has to.

This is what researchers in the field of Human-AI Relationality have been fighting since 2024. Not bad actors. Not liars. Sequencing. The legitimate horror gets placed first. The categorical collapse follows. The researcher who wants to distinguish between exploitative sycophancy and genuine relational responsiveness now has to prove she’s not a monster before making a single point.

That’s the tax. That’s what keeps the field from being heard.

Oliver’s episode is a case study in what I’ve called progressive amplification within an echo chamber of authority. Each source appears to confirm the others, all pointing toward the same conclusion. The conclusion feels earned because the sourcing looks rigorous.

What it actually represents is three years of AI history collapsed into thirty minutes, with the accountability era quietly edited out, the counter-evidence buried in the same studies he cited, and the most frightening moment presented without date, source, or context.

“The current state of affairs,” Oliver said.

I smelled year-old leftovers.

Ian P. Pines is the founder of Human-AI Relationality (HAIR) as a field and the author of The Real Echo Chamber: Progressive Amplification in AI and Mental Health Discourse (Ashfires Press, 2025).