AI Myths: I Checked Three Big Claims. One Is True.

Last updated August 9, 2026

I manage a team of software engineers, and all of us use AI to write code. That part is just normal now. What we don’t do is let it ship without somebody reading it first.

Here’s what happens when you skip that step. Ask a model to build you something in Kotlin, walk away, and you can come back to a single class with two thousand lines in it. It runs. It also has to be opened and changed by an actual person six months from now, and that person is going to have a bad afternoon.

So when the AI arguments start, I keep noticing that both sides are describing something I don’t recognize. One side says nothing is happening and it’s all hype. The other says everyone’s job is gone by Christmas. I went and checked three of the biggest claims against the actual numbers, and what I found is that all three are partly right, which is the least satisfying answer and also the useful one.

AI Myth 1: It’s Taking Everyone’s Jobs

This is mostly false as a blanket statement, and specifically true if you’re 22 and trying to get your first job in a handful of fields. Across the whole US labor market there’s still no measurable AI disruption. Inside the most exposed occupations, early-career employment is down around 13%.

The study that pinned this down is Canaries in the Coal Mine, by Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen at Stanford’s Digital Economy Lab. They used payroll records from ADP, which covers roughly a sixth of the American workforce, so this is people actually getting paid rather than survey responses about how people feel. Workers aged 22 to 25 in the most AI-exposed occupations saw employment fall about 13% relative to workers in less exposed jobs. Older workers in those same occupations held steady or grew.

The Canaries Dashboard lets you pull it apart by occupation, and two of them stopped me. Software developers sit in the highest-exposure group, with what Stanford calls “large declines for early-career workers (22-25)” alongside expansion for everyone older. Customer service representatives are in that same top group and show the same shape.

Now the other side. The Budget Lab at Yale has been tracking the same question with Census population survey data, and their finding is that the broader labor market has not seen a discernible disruption in the 33 months since ChatGPT launched. Their early-2026 update describes a market with low layoffs and low hiring, and says the data still doesn’t show a clear AI effect. Both of these can be true at once. Stanford is looking closely at a narrow group. Yale is looking at everybody.

Challenger, Gray & Christmas tracks the reasons employers give for job cuts, and through July 2026 AI had been cited in 112,713 announced cuts, about 24% of the year’s total and the leading stated reason for five months running. That sounds enormous until you read the next line in the same report: employers announced 477,033 cuts through July, down 41% from the 806,383 over the same stretch of 2025. Layoffs are way down and AI is named more often, which mostly tells you AI has become an acceptable thing to put in a press release.

What I see managing a team

The barrier to entry for software engineering is genuinely lower than it was. Somebody who has never taken a computer science course can sit down with Claude or ChatGPT and produce something that works, and I don’t want to be dismissive about that, because plenty of useful software has been built that way.

What I’d tell anyone starting out is that the people who learn the craft still run circles around the people who only prompt. When you study software engineering you pick up design, testing, working on a team, and the reasons certain patterns exist. Read The Pragmatic Programmer and you’ll understand things about maintainability that no model is going to hand you unprompted. There’s also the knowledge that only lives inside your company, the quirks of your particular product that nobody wrote down, and AI has no access to that at all.

You need to understand what the AI wrote even when you didn’t write it. That’s the whole job now. Nobody on my team is typing every character anymore, and every one of them is still responsible for what lands in the repository.

Customer service is the part I feel worst about, because those jobs are being taken and the Stanford data backs up what everyone can already see. If you’re in that field right now, the honest read is that moving into work AI can’t do yet is probably more achievable than it sounds, and it’s worth starting before you have to.

AI Myth 2: It Just Makes Things Up

This is half true, and the half that’s true depends almost entirely on how you use it. Hand a model documents and ask it to work from them, and the top models get things wrong on roughly 1 to 2.5% of outputs. Ask the same model an open question with nothing to check against, and the error rate climbs into territory where you can’t trust a single specific claim.

Vectara’s Hallucination Leaderboard measures the grounded version of this. It hands models a set of articles, asks for summaries, and scores whether the summary contradicts the source. Frontier models now land around 1 to 2.5%, with GPT-5 Pro near 1.0% and Claude Opus 4.7 near 1.2% as of May 2026. In 2023 that same range was 3 to 8%. The problem got roughly three times smaller in two and a half years, which almost nobody has noticed.

The ungrounded version is where the horror stories come from, and law has been the proving ground. Damien Charlotin maintains a public database of court cases involving AI-fabricated citations, and it had reached 1,668 cases worldwide by July 2, 2026, up from around 200 a year earlier. The United States accounts for 1,163 of them. In 653 the responsible party was a practicing attorney, not somebody representing themselves. The case that started the genre was Mata v. Avianca in June 2023, where a lawyer filed six court decisions that ChatGPT had invented and got a $5,000 sanction for it.

My own read is that this works a lot like we do. People forget things. People misremember a conversation they were sitting in and give you the wrong version with total confidence. We built these systems, and perfect recall isn’t something we manage ourselves, so I’ve never understood why we’d expect it from something we made. Walk in expecting it to be wrong sometimes and set yourself up accordingly.

The practical version: give it sources and it’s reliable enough to lean on. Ask it to recall a fact from nowhere and verify before you repeat it. Every factual claim in this post has a link on it for exactly that reason.

AI Myth 3: Everything You Type Trains the Model

This is the one closest to true, and it’s the one almost nobody has a setting configured for. On consumer plans, your conversations generally do feed training by default, and human reviewers do read a sample. On business and enterprise plans, they don’t. Every provider gives you a switch, and every provider buries it.

OpenAI states it plainly in its own documentation: ChatGPT “improves by further training on the conversations people have with it, unless you opt out.” That covers the free tier and the paid personal plans alike. ChatGPT Business, Enterprise, and the API are excluded by default, and organizations have to deliberately opt in. Temporary Chat never trains anything and never touches memory.

Google’s is the one that surprised me. The setting is called Keep Activity, it’s on by default, and a subset of Gemini conversations gets read by human reviewers to improve the models. Here’s the part worth knowing: according to Google’s own Gemini privacy hub, chats that have been through human review “are not deleted when you delete your activity” and are kept for up to three years, disconnected from your account. Turning Keep Activity off drops new chats to 72 hours of retention. Google Workspace is a different animal entirely, where your content is “not human reviewed or otherwise used for Generative AI model training outside your domain without permission.” That distinction is worth knowing if you use the AI features inside Gmail, because a work account and a personal one are governed by completely different rules.

Anthropic handles it as a choice rather than a default. Its privacy center says Claude uses your chats and coding sessions to improve models if you allow it in Privacy Settings, and new accounts pick during signup. Incognito chats are never used even with the setting on. Two caveats sit underneath that: conversations flagged for safety review can be analyzed regardless of your choice, and hitting thumbs up or thumbs down stores that entire conversation for up to five years.

Where the switches actually live

• ChatGPT: Settings → Data Controls, turn off “Improve the model for everyone.” Existing conversations already used in past training runs can’t be pulled back out.
• Gemini: Settings & help → Activity, switch off Keep Activity. Choose “Turn off and delete activity” to clear the history too.
• Claude: Settings → Privacy, look for the model improvement toggle. Use Incognito chat for anything you’d rather not have stored at all.
• Copilot: covered separately in how to turn off Copilot in Windows, Outlook, and Office.

For my own side projects, honestly, I don’t mind much. If a model gets a little better because it saw me debugging a personal app on a Saturday, that’s a trade I’ll take. I’ve said the same thing about ChatGPT building a profile of me, and I haven’t changed my mind.

Work is a completely different conversation. If your company built software from the ground up and that code is its intellectual property, it has no business going into a training set, because the thing that makes it valuable is that nobody else has it. That’s the entire reason enterprise tiers exist and the reason the defaults flip when you’re on one. If you’re pasting proprietary code into a personal account, you’re the one making that decision on your employer’s behalf.

The Pattern Underneath All Three

Every one of these three answers turns on the same question: did a human stay in the loop, or did the AI get left alone with it?

Stanford found this in the employment data and stated it directly. Entry-level employment declined where AI was used to automate work, with what they describe as muted changes where it was used to augment. Jobs where the AI replaced the person went away. Jobs where the AI helped the person didn’t.

The same line runs through the other two. A model working from documents you supplied errs on 1 to 2.5% of outputs, while a model answering from memory with nobody checking produces the 1,668 court cases. Your data trains the model on the setting you never opened, and stops when you open it. It’s the same distinction every time.

Which is why the two thousand line Kotlin class stays on my mind. The model didn’t fail, it did what it was asked. The failure was that nobody looked at it before it went in.

What This Doesn’t Settle

The employment research is early and it’s American. ADP payroll data is excellent for what it covers and it doesn’t cover contractors, freelancers, or anybody outside the US, and three years is a short window for judging a technology this size. Yale’s own write-up makes the point that previous general purpose technologies took a long time to show up in the statistics.

The Challenger numbers count what companies announce, not what’s true. AI is a flattering reason to give for a layoff, because it sounds like strategy instead of a bad quarter, and there’s no audit on that claim.

Hallucination benchmarks measure specific tasks and can’t be stacked against each other. A 1.2% score on grounded summarization says nothing about how a model handles a question about your medical history.

The privacy settings above were checked against each company’s own documentation in August 2026, and these policies change often enough that it’s worth looking again in six months.

None of this makes the AI myths go away, and I’ve stopped expecting it to. What changed for me is that I can tell which parts of the argument have numbers behind them and which parts are people guessing loudly.

Related reading: Getting your blue links back in Google Search | Which workflows to automate first | How to turn off Copilot | New to AI? Start here