Generative AI
Generative AI is artificial intelligence that produces new content — text, images, audio, video, code, structured data — rather than only classifying, scoring or predicting from what it is given: large language models generate text and code, diffusion models generate images and video. It became a mass technology with ChatGPT in late 2022, reached fintech as assistants, document processing and code generation — and, from the other side, as the synthetic documents, faces and voices fraud teams now defend against.
- Also known as
- GenAI, Generative artificial intelligence, Generative models
- Used at
- customer service and assistants, document extraction and summarisation, marketing and communications, code generation and software development, research and analysis, design and media, and synthetic data; and, adversarially, deepfakes, synthetic identities, phishing at scale and forged evidence
- Binding
- the EU AI Act's transparency obligations (AI-generated content must be marked; users must know they interact with AI; deepfakes disclosed), its general-purpose model obligations on providers, and high-risk rules where generative systems feed regulated decisions; copyright law on training data and outputs, under litigation; data protection on prompts and outputs; sector rules on what a generated communication to a customer may say
- Typical economics
- the labs' revenues grew from near zero in 2022 to tens of billions a year by 2025; enterprise adoption moved from pilots to production over 2024–2025; venture funding concentrated in AI; the largest transactions were infrastructure and licence-and-hire deals; application companies were valued on growth and diligenced on whether their product was more than a prompt
- Typical risks
- hallucination (confident falsehoods), inconsistency, prompt injection, data leakage through prompts, copyright exposure, bias in outputs, over-reliance by users, and the fraud uses; the mitigations are retrieval and grounding, verification, guardrails, evaluation, human review, and, on the defensive side, liveness and deepfake detection
- Parties
- the labs, the clouds and chip makers, the application builders, the enterprises deploying, the creative industries contesting training data, the regulators, the fraudsters, and the identity and trust-and-safety vendors
How generative AI works in practice
A fintech's operations team adopts generative AI in three places. Support: a language model drafts replies from the knowledge base and the customer's record, an agent reviews and sends — handling time down 40%, quality up, and a guardrail that prevents the model from promising anything about money. Documents: a vision-capable model extracts figures from statements and contracts with citations, verification recomputes them, and underwriters review — time per file down 70%. Engineering: coding assistants write and review code, tests and documentation — throughput up, with review discipline unchanged. Each use has an evaluation set, a human in the loop where the output matters, and a data-terms review of the vendor. None of them decides anything about a customer; the decisions stay with the models and people that were governed already.
The other side arrived as fast. Generated faces defeat naive selfie checks; injected synthetic video defeats liveness that was not built for it; generated documents pass template-based forgery detection; cloned voices pass voice authentication and phone verification; generated phishing at scale raises the base rate of every social-engineering attack. Identity vendors rebuilt liveness and injection detection; fraud teams added generated-content detection and consortium signals; voice authentication lost its standing as a sole factor. For a fintech, generative AI is productivity in operations and a new threat model in fraud, and both need a budget. For a buyer, a "generative AI" company is diligenced on what it built around the rented models — data, workflow, evaluation, distribution — and on the copyright and data-terms exposure of what it trained; an identity or fraud vendor is diligenced on how its detection has kept pace with the generators.
For a fintech founder, generative AI is a capability to deploy with governance and a threat to defend against with budget; for an investor, it is a value chain that rewards compute, models and distribution, and an application layer that must prove it is more than a prompt.
Example
A neobank sees two effects of generative AI in one year. Productivity: a support assistant on a rented language model with retrieval over its policies drafts 70% of replies, agents edit and send, handling time falls from 9 to 5 minutes, and CSAT rises — with an evaluation set of 1,000 conversations run on every change and a hard rule that the model never confirms a transaction or a balance without a system lookup. Threat: synthetic selfies injected into its onboarding flow rise from near zero to 1.5% of attempts; its identity vendor's upgraded liveness and injection detection catch 96%; the residual, traced through consortium data, reveals a synthetic-identity ring of 700 accounts. The bank's board sees both in one report: €2.1 million of annual cost saved, €0.9 million of fraud prevented, and a standing budget line for both.
Common mistakes
- Deploying generative outputs to customers without a human or a verification step where the content matters. - Vendors whose terms allow training on the company's prompts and documents. - Identity and fraud stacks built before generated faces and voices existed and never upgraded. - Reading a generative-AI application's growth as a moat when the model is rented. - Ignoring the transparency rules on AI-generated content and interactions.