Here is a story about a small AI team that shipped a model update and nearly lost two of its most valuable contributors without noticing. It is made up, but if you have ever worked on machine learning projects you will probably recognise every beat. The lesson is simple: when a team learns to credit every contributor for the specific thing they did, trust goes up, handovers get friendlier and the next release is easier to ship.
The release that went out with one name on it
Our imaginary team built an image-sorting service for a warehouse client. Four people did the real work. A data annotator relabelled a few thousand ambiguous photos by hand after spotting that two categories kept bleeding into each other. A platform engineer rewrote the training job so it stopped failing halfway through the night. A researcher on loan from another group suggested a different loss function. And a support analyst, who talked to the client every week, noticed that the model struggled with photos taken under yellow lighting.
The release notes said: “Improved classifier accuracy. Shipped by the ML team.” That was it. The annotator read it and felt invisible. The analyst read it and wondered why they had bothered passing on the lighting observation. The researcher, who was already half-committed elsewhere, quietly decided not to return for the next cycle.
Nobody was careless on purpose. The team lead simply wrote the notes on a Friday evening, tired and pleased, and used the shorthand that came to mind.
The moment somebody noticed
A week later the platform engineer, scrolling the repository history, noticed that the pull request description credited nobody. They mentioned it in the team chat, half joking: “we are very generous with each other’s work, aren’t we?” The room went quiet in the good way. The lead replied, “That’s on me. Let me fix it properly.”
What followed is the useful part. Instead of a vague apology, the lead wrote four short credit lines, each saying who did what and why it mattered:
- The annotator’s relabelling pass cleaned up the two confusable categories and gave the model clearer examples to learn from.
- The platform engineer’s rewrite of the training job made overnight runs finish reliably, which gave the team more experiments per week.
- The visiting researcher’s loss-function suggestion was the single change that moved the numbers on the hardest cases.
- The support analyst’s note about yellow lighting led to a new set of test photos that exposed a weakness nobody had measured.
The lead added those lines to the changelog, edited the pull request description, and posted a thank-you in the shared channel that quoted each contribution in the same words. It took about fifteen minutes.
What changed on the team afterwards
The shift was not dramatic, but it was visible. The annotator began volunteering notes about odd patterns in the data, because it was now clear those notes were read and valued. The researcher agreed to come back for the next cycle. The analyst started attaching a short “what the client is seeing” paragraph to each weekly summary, and those paragraphs fed directly into the test set.
The team also adopted a small habit worth copying. Every pull request template gained one extra field: “Who helped, and with what?” It is optional, but it prompts the author to think about the people behind the diff before pressing merge. Over a few months the field became the most-read part of the description, because it told reviewers whom to ask when something broke.
How to credit well when you work with AI systems
The same principle applies at every layer of an AI project, not only inside the immediate team. Here is a quick checklist you can adapt:
- Credit the data work. Labelling, cleaning and sampling decisions shape a model more than most people admit. Name who did them and what they changed.
- Credit the people who spotted failures. A user or support colleague who reports a strange result has done a piece of testing for you.
- Credit sources and prior work. If a technique came from a published paper or an open-source maintainer’s documentation, cite it where the technique is used.
- Credit the unglamorous infrastructure. Reliable pipelines are why experiments get run at all.
- Say it quickly. A thank-you written the same week carries more weight than a retrospective footnote.
If you want a refresher on the kind of work that benefits from this habit, our own pieces on how image classification systems look at pictures and on applying AI to tune up production lines both describe projects with many hands behind them.
Pointing readers to the original source
Crediting is not limited to people on your payroll. When your write-up leans on an article, a dataset or a talk, link to the original rather than to a summary. A handy everyday example is a news aggregator like HeadlineSift, which gathers stories and lets readers click each card through to where it was first published. The pattern, collect and then point onward, is the same one a good changelog follows.
Common slips in the story, and how to dodge them
Our fictional team made three mistakes that real teams repeat. The first was crediting the group instead of the person. “The ML team” is a label, not a thank-you. The second was delay: the correction arrived a week after the release, which is better than never but cost some goodwill. The third was assuming everybody knew. The annotator did not know the lead noticed the relabelling effort until the note appeared in writing.
The remedies are cheap. Keep a running list during the sprint of who helped with what, so release day becomes copy-and-polish. Use the same plain verbs each time: relabelled, rewrote, suggested, reported, tested. And make the credit visible in the places people actually read, namely the changelog, the pull request and the demo script.
Try it on your next demo
Before your next demo or release, pause for ten minutes and write one line per contributor. Say what they did and what it made possible. Read the lines aloud to a colleague; if any sounds like it could be pasted under a different project, make it more specific. Then publish them where the whole team and the original sources can be seen. The people behind your model will notice, and so will everyone thinking about joining your next project.