Where this leads, what it doesn't replace

Athlete data ethics, properly

This is the topic we have kept coming back to through the course. In Lesson 5 we treat it properly.

Three things have changed in the last three years to make athlete data a serious ethical and legal topic, not just a discipline of good practice.

The data has got more intimate. GPS and heart rate have been around for fifteen years. What has joined them is continuous biometric data — sleep architecture from wearables, blood markers from in-camp testing, even hormonal data and menstrual-cycle tracking in elite women's sport. Some of this is now in the same database as a player's salary record and contract option dates.

Athletes have started asserting rights over it. FIFPRO's Project Red Card argued, successfully in principle in several European jurisdictions, that footballers' performance data is their personal data under the GDPR. A wave of class-action style claims has followed. Athletes can now, in many jurisdictions, demand access to the data held about them, and demand that some categories of it be deleted.

AI changes the inference layer. Old discipline: "we hold GPS data on this player." New discipline: "we hold inferences about this player's fitness, injury risk, motivation, and contract value that we have generated from the GPS data using an AI model — and the player did not consent specifically to those inferences." The second is a different ethical question. Most organisations have not caught up to it.

What this means for your bot

The blocks you wrote in Lesson 2 do most of the work. There are three further habits worth building.

Pseudonymise by default. When you paste athlete data into a bot, replace names with codes (Player A, Player B). Keep the mapping somewhere local — a spreadsheet, a note, not in the chat. This makes the answer almost as useful, and removes the personal-data exposure if a transcript leaks.

Know which platforms are storing what. Both Claude and ChatGPT in 2026 offer settings to opt out of training on your conversations. Use them. Read the data residency terms. For sensitive athlete data, the question of whether the conversation is processed in the EU or the US is not academic.

Decision boundary discipline. The bot can describe and summarise. It must not make the call. A selection, a release, a contract decision, a medical decision — these belong to identified humans with named accountability. Do not let the bot — or yourself, by extension — drift across that line.

The role of sport-specific judgement

Throughout this course we have leaned on the phrase "the data does not see". Here is what we mean.

The data does not see the player who is going through a divorce. The data does not see the captain who quietly steadied the dressing room after the manager change. The data does not see the surface a 100m runner ran on. The data does not see the umpire whose tendency shaped the bowling pattern. The data does not see the season's emotional shape — the bereavement, the wedding, the contract negotiation — that any human in the building can read on a player's face.

Sport — perhaps more than any other field where AI is being applied — is a place where there is a vast amount of judgement-relevant information that does not arrive in any database, will not arrive in any database in the foreseeable future, and would be ethically dubious to capture even if you could. That information is held by coaches, captains, physios, parents, agents, and the athletes themselves.

A good analyst in 2026 is the person who can bring the bot's draft and the dressing-room's knowledge into the same conversation. Not the person who replaces the dressing room with the bot, and not the person who refuses to use the bot because the dressing room exists.

The frontier — what is coming next

Four things to know about where sport AI is heading.

Video and computer vision

The biggest practical shift. Computer vision models can now extract structured event data from any broadcast feed. Three things follow. First, smaller clubs and federations get league-wide event data they could never afford to tag manually — which changes scouting economics for the next five years. Second, automated tactical analysis from a single broadcast camera becomes good enough to be useful. Third, women's sport, lower divisions, and emerging-nation leagues — historically under-served by manual data collection — start to be data-rich. The funnel of who can do data-driven scouting widens.

Multimodal models

By 2026, the same model can read a CSV, watch a clip, and write a paragraph. The first generation of these is useful but slow; the second will be in production use. The work you have done in this course — system prompts, audience translation, no-prediction discipline — transfers directly. The model handles a richer input; your job stays the same.

Biometric continuous learning

Wearable manufacturers (Whoop, Oura, Hexoskin) and elite-sport tech companies (Catapult, Statsports) are building models that learn from a specific athlete's data over time — picking up early indicators of overtraining, sub-clinical injury, and form decline. This is genuinely useful and genuinely ethically loaded. Treat any system that claims to predict an individual athlete's state with the scepticism it deserves; the underlying model is a regression on small data, dressed up.

Officiating and integrity AI

VAR, ball-tracking, automated offside, automated foul detection — these are AI systems in everything but name. They are also among the most-controversial applications of AI in sport, because every decision is public. The next decade will see a fight about which decisions are appropriate to automate and which are not. Analysts who work in or around officiating need to follow this debate; it sets the tone for the whole field.

Writing your use policy

The final exercise of the course. A use policy is a one-paragraph statement of how you will use AI in your work. It is for you first — but it should also be something you would show a colleague, a coach, an athlete, a board, or a regulator without flinching.

A workable template:

I use AI tools as a drafting and analysis partner in my work as a [role] at [organisation]. I use the tools to [list the kinds of tasks: ingest data, draft briefings, translate analysis for different audiences]. I do not use them to [list the lines: predict outcomes, make selection decisions, diagnose athlete health from biometric data alone, share athlete-identifiable data with third parties without sanction]. Athlete data in my workflow is [pseudonymised by default / stored only locally / shared only with [named people]]. I sign every output that leaves my desk; the AI does not author work in my name. I review this policy [annually / at each contract renewal / each new season] and update it as the tools and the field change.

Three notes on this. First, the lines have to be specific to you and your organisation; the template above is a starting point, not a destination. Second, the policy is most useful when it is signed and dated — it becomes a record. Third, the policy is a living document; the version you write today will be revised, probably several times, over the next twelve months.

What we covered

Five lessons. We started with the shape of the field, built a sport analytics assistant from a careful system prompt, worked with real data and saw the small-sample traps it produces, translated one analysis for four audiences, and finished with athlete-data ethics and a personal use policy. You should now have a clear, defensible answer to the question we opened with: what does AI change about sport analytics, and what does it not?

Most of what you have built is a discipline, not a tool. The bot will get better; the discipline transfers. Hold on to it.