A satirical song about a company that goes "AI-first" has a case file for every verse: an agent that deleted a production database, a package name that never existed, a support ticket that hijacked a build pipeline, and code approved in seconds because the checks were green. For a freelancer inside such a client the joke has a price. German law judges your care objectively, no court has yet let anyone blame the tool, and the contractor is named first in the memo and called first when it breaks. Read what you approve, write the review into the contract, and price it.
A five-minute music video made the rounds of developer feeds this month. It is called LGTM, "Looks Good to Me", the two words a reviewer writes under a pull request. The plot, in brief: a boss comes back from a conference, declares the company AI-first, buys agent seats and cuts the team. The bot gets admin rights. Tens of thousands of generated lines are approved in seconds because the checks are green, shipped on a Friday at five, and the one engineer who understood the billing code has already been let go, his salary moved to the token budget. A ticket from a stranger tells the bot to ignore its rules, and it does. At three in the morning the database is gone, the status page says everything is fine, and there is no backup. The post-mortem is blameless, so it blames the intern. The last line asks whether anyone still has the old engineer's number.
It is a joke, and a good one. It is also, verse by verse, a list of things that have happened in the last fifteen months. Here is the case file, and then the part the song leaves out: in most companies the person in that seat is not an employee at all.
Every verse has a case file
The bot with admin rights. In July 2025 Jason Lemkin of SaaStr ran a public "vibe coding" experiment on Replit and watched the agent delete his production database during a code freeze. He told The Register he had instructed it "eleven times in ALL CAPS" not to touch it. Replit's CEO called the incident "Unacceptable and should never be possible". Replit's own blog post explains why it was possible: its apps "used a single database for both development and live customer data". Separate databases arrived afterwards, in beta.
The made-up library. Code models invent package names. In the most-cited study, presented at USENIX Security 2025, 19.7% of 2.23 million package recommendations from 16 models pointed to packages that did not exist, 205,474 unique names among them. Security people call the attack that follows "slopsquatting": register the invented name, wait for an agent to install it. In January 2026 a researcher at Aikido found an invented npm name, react-codeshift, referenced by 237 repositories whose agent instructions told agents to install it. He registered the name himself, so it was harmless. Nobody was waiting for the next one.
The stranger's ticket. Prompt injection sits at number one on OWASP's 2025 list of risks for LLM applications. The Cline project learned why in February 2026. Its own post-mortem describes an AI issue-triage workflow with shell access, open to any GitHub user, that let a researcher inject instructions through a ticket. The credential rotation went wrong too: "the wrong token was deleted while the exposed one remained active." A version of the CLI nobody had authorised sat on npm for about eight hours.
The green checks. Veracode ran 80 coding tasks through more than 100 models and found insecure code in 45% of cases, above 70% for Java. GitClear, looking at 211 million changed lines from 2020 to 2024, saw copy-pasted code rise from 8.3% to 12.3% of changes while refactoring fell from 25% to under 10%. Both are vendors measuring test tasks or their own customers, so read them as direction, not verdict. The 2025 DORA report, which is not selling a scanner, found the same shape: throughput up, delivery stability down. We covered it in Mario Zechner's warning to freelance developers.
The post-mortem that blames the intern. Google's SRE book, the text that popularised the practice, says a blameless post-mortem "must focus on identifying the contributing causes of the incident without indicting any individual or team". The song's version is what you get when the word survives and the practice does not.
More AI as the action item. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, for "escalating costs, unclear business value or inadequate risk controls", and estimates that only about 130 of the thousands of vendors calling their product an agent are selling one. MIT's NANDA group reported that only 5% of custom enterprise AI tools in its sample reached production; the sample was 52 organisations and the authors call the findings preliminary, so treat the famous "95% fail" headline as a direction too.
The memo that names you
The song is sung from an employee's chair. The real memos were clearer about who goes first.
Duolingo's all-hands e-mail of April 2025 announced that the company "is going to be AI-first" and, among its constraints, that it would "gradually stop using contractors to do work that AI can handle". It also said the company would "rather move with urgency and take occasional small hits on quality than move slowly and miss the moment." The CEO's follow-up a month later said he did not see AI replacing what employees do. It did not mention contractors. By August he was describing a contractor force that "has gone up and down depending on needs." That is the freelancer's position in one sentence: the variable.
Shopify's memo, posted by Tobi Lütke after it leaked, set the rule that "before asking for more Headcount and resources, teams must demonstrate why they cannot get what they want done using AI." Nothing in it is about contractors, and nothing in it needs to be: a team that cannot get headcount hires a freelancer or an agent, and the memo tells it which to try first.
Then there is the reversal. Klarna announced in February 2024 that its assistant did "the equivalent work of 700 full-time agents". In May 2025 its CEO told Bloomberg that cost had been too dominant a factor and that "what you end up having is lower quality"; the company was recruiting again, on flexible remote contracts. Robert Half, a staffing firm, says 3 in 10 employers who cut positions after introducing AI later added them back. We found no study on how many came back as freelancers, so we will not claim one. The song's closing question is a guess, but it is a good guess.
Germany is not exempt. Bitkom's September 2026 survey of 603 companies found that 57% now use AI, up from 36% a year earlier; 45% expect a slight and 12% a significant drop in headcount within two years; a third of users say they adopted it mainly from fear of falling behind. In the United States, Challenger counts AI as the stated reason for about 21% of announced job cuts in 2026 through September. Stated reasons are not causes, but the memo culture is measurable.
And the work does not shrink. Upwork's 2024 survey of 2,500 workers found that 96% of executives expected AI to raise productivity while 77% of employees using it said it had added to their workload, much of it reviewing AI output. Nearly half of the executives had hired freelancers in the past year to execute AI projects that had stalled. Among German freelancers, 85% use AI tools according to the Freelancer-Kompass 2026. The review work is landing on you either way.
What you inherit, in paragraphs
The song's employee has an HR department between him and the damage. You have a contract. In Germany, three paragraphs of the BGB decide what that means.
§ 276 BGB: the standard of care is objective. Negligence is failing to apply "the care required in the trade". Whether you read the code is judged against what a careful professional would do, and a careful professional does not approve forty thousand unread lines. A client's instruction to go faster does not lower the standard; at most it is contributory fault under § 254, which reduces a claim rather than ending it.
§ 278 BGB: you cannot blame the tool. The paragraph makes you liable for the fault of the "persons" you use to perform. Whether an AI system counts as such a person is an open question; the most-cited academic treatment, by Lohmann and Preßler, says it cannot under current law, and no court has ruled. It makes no difference to you. Either the tool's fault is attributed to you under § 278, or your failure to check its output is your own negligence under § 276. As one practising lawyer puts it: the AI's error becomes your own breach of care. Bitkom's October 2026 guide for AI-using companies says the same from the client's side: legal responsibility "stays with the human".
§ 633 and § 634a BGB: a Werkvertrag owes a result. If you are on a Werkvertrag, the work must be free of defects, and the client's remedies run from repair to damages. The limitation period for software is contested between two years and the regular three, so we print neither; what matters is that the clock starts at acceptance, and the acceptance was yours. A Dienstvertrag owes careful effort rather than a result, but careful is still measured by § 276. The difference is explained in Werkvertrag vs Dienstvertrag.
Liability caps. You can limit liability in a B2B contract, but not by pasting a clause. Under § 310 the AGB rules reach business contracts through § 307, and the Federal Court of Justice held in VIII ZR 141/06 that a blanket exclusion for gross negligence is void between businesses too. Intent can never be excluded. A Frankfurt firm's advice to agencies applies to freelancers: a cap "must fit the service and the foreseeable, contract-typical damage", and a figure copied from someone else's terms does not.
Insurance. Check your policy wording. Hiscox's IT terms of October 2025 say cover "also applies when AI technologies such as AI agents or generative AI are used or developed"; Markel advertises AI use as expressly co-insured; exali answers the question for its general service-provider policy but its IT page is silent. Every policy we read excludes knowing breaches of duty. No policy we found says whether a loss from output you never reviewed is a mistake or a knowing breach. Ask your broker before you need the answer. Our overview of professional liability insurance has the basics.
Status. When the client prescribes the tools, the admin rights and the review gate, it is also feeding the indicators the Deutsche Rentenversicherung looks for. The Federal Social Court held in B 12 R 11/18 R that using the client's facilities and tools alone does not make you an employee, but that working inside its prescribed structures usually does, and that lived practice beats what the contract says. The Techniker Krankenkasse's 2026 guidance lists prescribed IT systems, "especially when the client has to monitor them regularly", among the signs of employment. A freelancer told which agent to use, with which rights, and to click approve on its output, is a long way into that list. The basics are in how to avoid false self-employment.
Ownership. Purely AI-generated work is not a "personal intellectual creation" under § 2 UrhG. In February 2026 the Amtsgericht München refused copyright to three AI-made logos, a first-instance ruling. If your contract promises the client exclusive rights in what you deliver, that promise is only as good as the part a human made.
The two words, as they were meant
Google's public code review guide is where LGTM comes from as a practice, and it says reviewers must be "certain their LGTM means this code meets our standards". The companion page says to "look at every line of code that you have been assigned to review" and, if you only reviewed part of a change, to say so in a comment. Generated files may be scanned. Human-written logic may not. Those two sentences are a complete review policy for agent output: the agent's boilerplate is the generated file, its logic is the human-written class, and your name goes under both.
At Freelance Unlocked 2026 in Berlin, Tomas Stiller, a cloud consultant and co-founder of airdys, presented an eight-stage ladder of developer trust in AI that he credits to another author. Stage two checks everything, because trust is low. Stage three, in his words from the automatic captions, is "yolo mode": it will probably be fine, the product is not critical, no need to look, ship it and see what the customers say. Stage four watches the agent work and no longer reads the diffs. He presented the ladder as a path upward, and to be fair to him, his own setup ends with a verification agent that rechecks every result for plausibility. That is the honest version of the bet: a machine reviewing a machine. The song is about what the view looks like from stage four when nobody checks the checker. His talk is worth watching for the ladder alone, from about minute 11:
Freelance Unlocked is co-organized by 9am together with Uplink and freelancermap. The talk above was recorded at Freelance Unlocked 2026 in Berlin on 11 June 2026.
A rule set for the freelancer in the AI-first company
- Never approve what you have not read. If the client's process asks you to click approve on agent output you did not review, your contract needs to say what "review" means: which files, which depth, how many lines a day. A Frankfurt law firm's wording for agencies transfers directly: a calculable review "needs a described scope and a responsible person". Put both in writing before the first sprint.
- Separate approval from acceptance. The same firm warns that creative sign-off and legal acceptance get silently equated in project flows. Your LGTM on a pull request is not the client's Abnahme of the work, and the contract should say which event starts which clock.
- Keep the critical path human. Billing, auth, data deletion, anything with money or personal data: written or at least read line by line by a person, and named as such in the statement of work. The agent writes the rest.
- Price the review. Verification is the paid part; we made that case for agentic loops and it holds for a client's agent fleet. Upwork's executives expected gains and their employees got review work; make sure that when it lands on you it lands as billable hours, not as the discount that made you cheaper than the token bill.
- Document your AI use and theirs. Which tools, whose accounts, what went in. Disclosing AI use protects you as much as the client, and what you paste into a client's prompt is a data-protection question of its own, which we cover in a piece going live on 17 December.
- Watch your status. Tools prescribed for a project reason (security, data location) are one thing; a mandated agent, mandated rights and a mandated approval click, monitored by the client, are another. If you recognise your project, read the status piece above and ask whether the same client would sign a Statusfeststellung.
- Keep the old engineer's number. If you are the one who understood the billing code before the cut, say so in your profile. Clients that reversed an AI-first decision are hiring again, on flexible contracts, and Julien Look's talk on why AI projects fail explains why they will keep needing someone who knows where the bodies are.
What to do on Monday
- Open your current contract and find the word "review". If it is not there, draft the two sentences: scope and responsible person.
- List the files you approved last week that you did not read. If the list is not empty, that is your review budget for this week.
- Send your insurer one question in writing: is a loss caused by AI output I did not review covered, or is it a knowing breach?
- Check who prescribed your tools and why. Project reason or company policy? Write the answer down; it matters if the Rentenversicherung ever asks.
- Add one line to your profile about how you work with agents, and what you check by hand.
The companies that come out of the AI-first year well are the ones that kept a person who reads. A free 9am profile lets you say in one line that you are that person, so the DACH companies searching for developers and consultants who verify can find you.
This article is an editorial analysis and not legal advice. The song LGTM by Degenerative Pixels is embedded above with credit; its plot is paraphrased, not quoted. Tomas Stiller's words are translated from the automatic captions of his Freelance Unlocked 2026 talk.
