I invent industry-defining approaches, and build the institutions that make them last.

Most media strategy still assumes the future is a mix: some broadcast, some internet; some human curation, some algorithmic; some direct relationships with audiences, some through platforms. I think that is wrong, and that the honest position is three hypotheses held together.

The first is that delivery becomes internet-only. Not internet-mostly. Every terrestrial, satellite and cable network is now a legacy asset with a closing date, and the only question is how well the transition is managed for the people who depend on it. The second is that interaction becomes AI-driven: the interface between a person and a body of content will be a conversation with a model, not a grid of thumbnails. The third is that platforms intermediate: between the institution that makes something and the person who experiences it sits an intermediary the institution does not control.

From these I derived five forces I use to test any plan, the "five i's": internet-only, intelligent, intermediated, interactive, immersive. A plan that does not have an answer for all five is a plan for a world that is already ending.

The consequence for a public institution is uncomfortable but clear. If you do not own your distribution, you must own your standards; if you do not own the interface, you must own the provenance of what flows through it; and if you cannot be the platform, you must be indispensable to every platform. That is the logic behind internet-first architecture, behind C2PA, and behind treating R&D as the place where the institution's future options are manufactured.

Every few months a new detector promises to identify AI-generated media, and every few months generation improves enough to defeat it. This is not a temporary state; it is structural. The generator and the detector are trained against each other, and the generator has the larger budget.

The mistake is to frame the problem as "is this fake?" The tractable question is "where did this come from, and has it been changed since?" That is a provenance question, and provenance is a solved problem in other domains: it is what a chain of custody is, what a signed certificate is, what a ledger is. C2PA applies that logic to media. A camera, an editing tool or a model signs what it produces; each subsequent change adds to the record; the viewer can inspect the chain.

Three things follow. Provenance has to be open, because a proprietary chain of custody is just another platform. It has to be adopted by the producers of media, which means the camera makers, the creative tools and the AI models, which is why the coalition had to include the companies that make cameras, creative tools and models. And it has to be invisible to work: nobody checks a certificate on a bank's website, they just notice when the padlock is missing. The goal for Content Credentials is the same. The measure of success is the day a piece of media without them looks suspicious.

The wider lesson for any regulated industry is that trust cannot be added at the end. It has to be designed into the supply chain of whatever you produce, and the organisations that do this first set the standard everyone else must meet.

Ask a finance director what R&D is worth and you will hear a cost. Ask a research director and you will hear a mission. Neither answer survives a difficult budget round, and that is why research is so often the first thing cut and the last thing rebuilt.

The alternative is to run the function the way a good investor runs a portfolio. Most bets fail cheaply; a few return many times their cost; the skill is in the allocation, the gating and the honesty about which stage each bet is in. At BBC R&D that meant an explicit pipeline, a patent portfolio managed as an asset rather than a trophy cabinet, external co-investment so the institution was never the only backer of its own future, and an independent assessment of the return.

That last point matters most. The DotEcon assessment put the net benefit of the function at between £5 and £9 for every pound invested. The number is useful; the fact that it was independent is what made it defensible. A research organisation that commissions its own audit before anyone demands one has changed its relationship with the people who fund it.

None of this is specific to broadcasting. Any organisation with a research, innovation or venture function can ask the same three questions: what is the pipeline, what is the return, and who else is investing alongside us?

The Partnership on AI brings together the laboratories building frontier models, the academics studying them, the civil-society organisations affected by them and, increasingly, the governments trying to regulate them. Chairing its board has taught me what such a body is for and what it is not.

It cannot regulate. Members compete, and no voluntary body can bind them. What it can do is three things governments and companies cannot do alone. It can establish shared vocabulary and evidence before regulation is written, so that regulation is written well. It can create the practical frameworks, on synthetic media, on safe deployment, on labour, that members adopt because adopting them is cheaper than each inventing their own. And it can be the place where a laboratory hears from the people its systems affect, in a setting that is not a hearing or a lawsuit.

The Chair's job is agenda and credibility: choosing the few questions that matter enough for members with different interests to fund, and representing the answers to the G7, the OECD and the national AI safety institutes in a way they can act on. It is the same discipline as convening a standard: give away enough that everyone can join, hold onto enough that the work has direction.

For any board, the transferable lesson is that governance of a technology nobody fully controls is a coalition problem before it is a compliance problem.