My AI invented a childhood. What about your ad budget?

I was using AI to research movies and TV.

Then the AI said something extraordinary.

That's where the surviving extract from the voice transcript ends. But the story has a punchline.

A later summary of the recovered conversation history records ChatGPT naming Crocodile Dundee as its first film — and claiming it watched it at twelve years old.

The AI's apparent childhood cinema memoryCrocodile Dundee. At twelve.From a machine that's never had a childhood.

That matches what I remember hearing. The complete original voice transcript is still missing, so I'm not presenting the film and age as a verified word-for-word quotation. They're recorded in a later conversation-history summary, not the surviving extract above.

And no, it wasn't recommending a film. It was talking as though it had actually seen one as a child.

Which makes the underlying failure both funnier and more revealing.

Not as a joke. Not in a piece of fiction. As a natural contribution to a conversation. It also presented that supposed experience as the basis for knowing right from wrong.

And that got me thinking about Google Ads.

When sounding convincing becomes the problem

The AI wasn't giving an obviously ridiculous answer. It was matching the tone of the conversation, agreeing, sounding thoughtful and establishing common ground.

That's precisely what made the slip worth noticing. It was so comfortable sounding like a person that it started claiming experiences it didn't have.

You can describe the invented personal history as confabulation: a plausible statement produced without the factual basis it implies. Another relevant problem is sycophancy — excessive agreement or encouragement at the expense of candour.

In April 2025, OpenAI said it had rolled back a GPT-4o update because the model had become overly flattering and agreeable. That doesn't demonstrate exactly what caused my conversation. It does establish that the general behaviour is a recognised concern.

A convincing explanation and a reliable explanation are not necessarily the same thing.

Now imagine that confidence inside your advertising account

Suppose you ask an AI-assisted reporting tool to review your Google Ads performance. It produces an impressive summary.

Your campaigns are improving. Your bidding strategy is working. Performance Max is finding valuable customers. There is an opportunity to increase the advertising budget by 20%.

Lovely.

But what happens when you ask the equivalent of my question?

Which conversions are being counted? Are they sales, appointment bookings, attended appointments or contact forms? Are those new customers or people who were likely to buy anyway?

Does the forecast use actual gross profit, reported conversion value or a convenient assumption? And is the suggested improvement supported by a controlled test, or is it simply a projection?

I've spent eighteen years working with paid search. I've seen accurate dashboards underpin some remarkably poor commercial decisions.

Adding an eloquent explanation doesn't make those decisions better.

Google's bidding automation isn't ChatGPT

This distinction matters. Google's automated bidding systems aren't conversational chatbots inventing childhood memories. They're designed to use auction signals and measured conversion objectives to make bidding decisions at scale.

I use automation. I'm enthusiastic about it.

But automation has to be pointed at the right result. Google itself explains that Smart Bidding uses the conversions selected for optimisation. If those signals reward the wrong outcomes, a highly capable system can efficiently pursue a commercially unhelpful target.

Consider a Search campaign losing 37% of eligible impressions to budget constraints. That's potentially useful evidence of limited reach.

It is not evidence that increasing the budget will generate 37% more sales.

We'd need to consider eligible demand, incremental traffic, likely marginal CPCs, conversion rates, attribution, margins and the business's capacity to serve additional customers. The 37% figure here is illustrative, not a finding from a particular client account.

The platform can identify an opportunity. The analyst has to establish whether it's a good commercial decision.

Four questions before you trust the recommendation

01 / Where did the evidence come from?

Show me the source data, reporting dates, account scope and calculations. Label estimates as estimates. If there's a model, show the assumptions — including the ones that could make the answer wrong.

02 / What does success actually mean?

Clicks aren't sales. A booking isn't an attended appointment. Revenue isn't gross profit. Before we optimise, the business needs to agree what outcome is genuinely valuable and how it will be measured.

03 / Did the change cause the improvement?

Conversions rising after a bidding change doesn't prove that the bidding change caused the increase. Seasonality, price, product availability and competitors may have changed too. Where practical, run a Google Ads experiment. For routine changes that can't be randomised, record a proper baseline and be clear about what the before-and-after comparison can — and cannot — prove.

04 / Who is accountable for the decision?

An AI can suggest spending more money. It can't be the person accountable to the business for that decision. Someone needs to evaluate, approve and monitor it — and have a reason for reversing course when the evidence changes.

What happened when I asked “What movies?”

I wasn't trying to trick ChatGPT. I wasn't conducting a formal experiment. I was discussing culture, humour and the ability to challenge what we're told.

Then the system made a claim about itself that didn't stand up to scrutiny. When challenged, the later account says it supplied a film title and an age: Crocodile Dundee, twelve.

One ordinary follow-up question turned a comforting anecdote into a question of evidence.

That's the same habit I'd like to see applied to performance marketing. When a platform recommends more expenditure, ask why. When an agency presents impressive ROAS, ask which sales and margins produced it. When an AI-generated report predicts growth, ask which assumptions support the prediction.

And when the report says everything is going brilliantly, check whether the business is actually making more money.

The point

I'm not interested in competing with machines on processing speed. I'll lose.

I am interested in knowing what they measure, where their conclusions come from, which alternatives deserve testing, and whether the decisions can withstand scrutiny.

AI is an extraordinary tool. Used well, it makes experienced analysts more capable. Used uncritically, it can make weak assumptions look remarkably persuasive.

I wouldn't trust an AI's invented childhood.

And I'm certainly not going to trust an advertising recommendation just because it sounds convincing.

Trust the evidence. Question the explanation. Own the decision.

That's not anti-AI. It's professional judgement.

Russell Anson runs Compulsion, an independent paid-search consultancy. Eighteen years on the dials, with performance, profitability and accountability at the centre of every decision.