Workers are rewriting their LinkedIn histories, and 2022 now looks 30% more AI-skilled than it was

A working paper analysing 29.4 million US LinkedIn profiles found nearly a fifth had retroactively edited a job they had already left, with AI-related additions rising more than sixfold since ChatGPT launched. The authors estimate a 2026 snapshot overstates 2022 AI skills by about 30%.

The record of who did what at work is quietly being rewritten. A National Bureau of Economic Research working paper tracking 29.4 million US LinkedIn profiles found that nearly a fifth had retroactively changed the title or description of a job the person had already left.

What gets added has a clear direction. Retroactive additions of terms including “AI“, “GPT“, “LLM” and “artificial intelligence” to previous roles have risen more than sixfold since ChatGPT was released in late 2022.

Things are also disappearing. By the end of 2025 workers were adding remote-work language about as often as they were deleting it, and additions of diversity and inclusion terms fell sharply in early 2025.

The authors call the practice time travel. The median edit landed more than four years after the job in question had ended.

It is not evenly spread. Technology and information leads at 31.6%, ahead of arts and entertainment at 25% and professional and scientific services at 24.1%, against a study-wide rate of 19.7%.

The researchers are careful about what this means, and so should everyone else be. Adding AI language is not proof of exaggeration, since people may be supplying detail they left out or describing old work in current vocabulary.

The finding that travels furthest is not about workers at all. The paper estimates that a 2026 snapshot of LinkedIn would overstate how common AI-related skills were in 2022 by around 30%.

That is a measurement problem for everyone. Career history is the substrate under a great deal of labour-market research and under the models sorting applicants, in a market where an AI-on-AI war already shapes hiring.

In Europe it became a compliance question three weeks ago. AI systems used in recruitment and employment decisions moved into the AI Act’s high-risk tier on 2 August.

Article 10 sets the bar those systems must clear. Training data must be relevant, sufficiently representative and, to the best extent possible, free of errors and complete, in a regime still arguing about jobs and how far the rules should reach.

Nobody drafted that clause with profile edits in mind. But a company training a hiring model on career history now has a published, numbered estimate of how much of that history was written afterwards.