In brief
Every AI system is a claim about its data. Train a model on what a business already knows, and it will reproduce that knowledge faithfully, including the parts the business would rather not repeat.
Every AI system is a claim about its data. Train a model on what a business already knows, and it will reproduce that knowledge faithfully, including the parts the business would rather not repeat. This is what makes data both the asset that makes AI valuable and the single largest source of legal exposure in the project. For a General Counsel, the useful move is to stop thinking of data risk as one problem and start treating it as five, because each fails in a different way and each is mitigated differently.
Privacy and data protection. Models are hungry, and the easiest data to feed them is data about people: customers, employees, applicants. Use it without a lawful basis and the company has not built a product, it has built a breach. The GDPR, California's CCPA, and India's DPDP Act 2023 (now operational through the DPDP Rules, 2025) each turn on consent, purpose limitation, and security, and they reach the training pipeline as much as the customer database. There is also a failure mode unique to AI: a model that memorises fragments of its training set and later repeats them to whoever asks. Privacy by design is the answer that scales, that is, minimise what you collect, anonymise or pseudonymise where you can, and run a privacy impact assessment on anything high-stakes before the first epoch of training, not after the regulator calls.
Bias that hardens into discrimination. A model trained on the past will defend the past. If historical hiring favoured one profile, or historical lending penalised one neighbourhood, the model learns that pattern and applies it at scale and at speed, with the added danger that it looks objective while doing so. The exposure is not only reputational; skewed outcomes in hiring, credit, or insurance can cross into anti-discrimination law. Bias rarely announces itself, because it hides in unrepresentative samples and in proxies that stand in for protected characteristics. The discipline that catches it is unglamorous: curate representative data, audit outputs against fairness metrics, and keep a human review in the decisions that change a person's life.
Quality, or garbage in, judgment out. Bias is the risk everyone discusses. Plain bad data is the one that quietly does the most damage. Missing fields, duplicates, stale records, and collection errors do not stop a model from producing an answer; they only stop it from producing a correct one, and a confident wrong answer is worse than no answer at all. A demand forecast built on dirty data will still forecast. Treat data as an asset with owners and lineage, clean and verify before training, and vet any third-party dataset for both accuracy and the right to use it, which leads directly to the next risk.
Intellectual property and ownership. Training data comes from somewhere, and somewhere usually has rights. Using copyrighted material to train a model, or scraping a site in breach of its terms, invites infringement and contract claims, and generative systems raise the stakes by producing outputs that can resemble the works they learned from. The ownership question runs the other way too and is genuinely unsettled: an invention devised by AI cannot, in most jurisdictions, name the AI as inventor, and content generated without a human author may attract no copyright at all. The practical safeguards are contractual and procedural, that is, license training data properly, read the IP and usage terms of any third-party model or API, forbid staff from feeding proprietary material into external tools, and make bespoke-development contracts say plainly who owns the model and its outputs.
Security of the data and of the model. An AI system widens the attack surface in ways a conventional application does not. Beyond ordinary breach, the model itself becomes a target: it can be stolen outright, poisoned during training so that it learns the wrong lesson, or fooled at inference by inputs crafted to be misread. A poisoned spam filter learns that spam is safe; an adversarial image walks past a vision system. Treat training data and model weights as crown-jewel assets, fold them into the enterprise security programme rather than a side project, restrict and monitor access, and write an incident-response plan that knows what to do when the thing compromised is a model, including how fast you can retrain and what you tell the people affected.
These risks are not a list; they are a web. Bias grows out of data quality and lands in discrimination law. A security breach becomes a privacy violation. The organisations that manage this well are the ones where the data scientists, the risk owners, and the legal team are looking at the same picture rather than three views of it. With the data risks understood, the next decision is structural, and often the most consequential one a company makes here: whether to buy an AI solution or build one.