M3GAN 2.0 Came to Life on the Street. The Hard Part Was Making It Safe.

A screen at Westfield Stratford City clocks the green jacket in the crowd. M3GAN 2.0 turns, looks straight at the person wearing it, and says “Hey you, in green.” Real-time. Personal. A little unsettling, which for this film was the point. Built with Universal Pictures and Ocean Outdoor, on a public screen, reacting to whoever walked past.

The reaction is what people noticed. What made it possible is what they did not.

A live AI character on a public screen, in front of families, reacting to strangers, is a legal and brand-safety problem before it is a marketing one. A film about an AI that goes dangerously off the rails cannot be promoted by an AI that goes off the rails. The reason this campaign shipped, and the reason a major studio put its name on it, is the part that never makes the highlight reel. It is also where the way we build everything since began.

A studio’s legal bar set the design

Universal Pictures did not approve a clever idea. They approved a system that cleared a high legal and compliance bar, and that bar shaped how we built the technology rather than the other way round.

That is the test most AI activations never have to pass, and exactly the one that catches them later. Gartner now predicts that by 2027, governance issues will push 40% of enterprises to demote or decommission their AI agents, gaps found only after something goes wrong in production. We built to clear the bar before launch, because for a campaign this visible there is no after.

So we built our own AI model

You cannot govern a model you do not own. You can wrap a policy around a rented model and hope it holds. You cannot gate it.

So we built our own. The character worked inside a fixed framework with no open internet to reach into and repeat a rumour, name a competitor, or say something the studio would never sign off. The questions from the public could go anywhere. The character could not. That boundary was built into the model itself, not bolted on top as a filter catching bad answers after the fact. It is the difference between a brand asset and a brand incident with a friendly voice.

It reacted to everyone and recorded no one

This is the line that mattered most to the lawyers, and it should matter most to any brand watching. When the industry says “brand-safe advertising” it usually means where the ad is placed, not running next to harmful content. We mean something different and harder: the AI itself is safe. The character reacts to a public audience and identifies, records, and stores no one.

The screen read what it needed in the moment to respond, and held on to nothing. No individual was identified, recorded, or stored. That mattered above all for children, who are the last people any brand wants sitting on a server somewhere. Sensing is not the same as surveilling, and the difference is a design decision you make on purpose, before a single person walks past.

So the experience felt like it knew you for one delightful second, and forgot you completely the next. Alive and safe at the same time.

The compliance behind the camera

Universal’s legal team did not take “it is safe” on trust. The camera recognition had to be demonstrably compliant, and that is what we built.

The character inferred only what it needed in the moment to choose a response. It ran on the device itself, so no image ever left the screen. It used no facial recognition and built no biometric profile. Everything was processed in the moment and discarded. Nothing was recorded and nothing was stored, so there was no personal data to hold, because no individual was ever identified.

That approach was designed to meet UK GDPR and the Data Protection Act 2018, backed by a formal Data Protection Impact Assessment and particular care for children’s data. Universal’s own legal and compliance teams reviewed it and signed it off before the screen went live.

Compliance was not a document produced after the fact. It was the architecture of the system.

The campaign that became our blueprint

We did not set out to write a rulebook. We set out to make M3GAN talk to a street. But the decisions that campaign forced on us, build your own model, gate it so it cannot wander, sense without storing, clear the legal bar before launch rather than after, turned out to be the decisions that matter on every screen, in every venue, for every brand.

So they stopped being one campaign’s requirements and became how we build the venue intelligence platform. M3GAN 2.0 is where the Performit Live approach was forged. Everything we have shipped since runs on what that project taught us: the character is the easy part, and the trust is the product. We did not add governance to our platform later. We learned it here, under a studio’s scrutiny, and built the platform around it.

This is the real future of AI advertising

Not that a screen can react. Plenty can now. The future is that it can react and stay safe, on a brand’s most visible surface, in front of the public, under a legal team’s scrutiny.

Reactive and governed. Most vendors can do the first. The first without the second is the thing Gartner is counting, and the thing that gets pulled off the wall after the launch photos. We treat the reaction as the easy part and the trust as the product, because that is what lets a brand leave the character running.

Watch how M3GAN 2.0 reacted live to East London shoppers using AI audience detection.

Want a character like this in the real world?

If you are weighing an AI character for a campaign or a venue, it is worth knowing the questions to ask before you choose. The honest first question is not whether it will turn heads. It is whether it will hold up under your brand’s, and your lawyers’, scrutiny.

Two minutes, no sales call: the AI Character Readiness Scorecard scores you on exactly that → performitlive.com/scorecard/. Or book a demo to see a character running live.

Brand-safe AI advertising is not about where an ad is placed. It is about the AI itself: an interactive character that reacts to an audience in real time without identifying, recording, or storing any individual, and that works inside a gated model it cannot leave, so it never says anything the brand would not approve.

FAQs

What was the M3GAN 2.0 AI advertising campaign?

An interactive AI character built by Performit Live with Universal Pictures and Ocean Outdoor at Westfield Stratford City, which reacted to passers-by in real time, for example “Hey you, in green”, to promote the film.

Where did the M3GAN 2.0 activation run?

On an Ocean Outdoor screen at Westfield Stratford City, East London.

Did the campaign record or identify people?

No. It ran on-device with no facial recognition. It sensed what it needed in the moment and stored nothing. No individual was identified, recorded, or kept, which mattered most for children.

How was the camera recognition compliant?

It ran on the device with no image leaving the screen, used no biometric profiling, processed and discarded data in the moment, was designed to meet UK GDPR and the Data Protection Act 2018 with a formal Data Protection Impact Assessment, and was signed off by Universal’s legal and compliance teams before launch.

How do you stop an AI advertising character from going off-script?

By building the boundary into the AI model itself rather than bolting a filter on top. The character works inside a fixed framework with no open internet to reach into, so it cannot wander off.

Why does an AI advertising character need its own model?

You cannot fully gate a model you do not own. Owning the engine is what lets the character be locked to a framework it cannot leave, which is what makes it safe enough for a brand to put its name on.

Can interactive AI advertising be brand-safe and privacy-compliant?

Yes. The M3GAN 2.0 campaign reacted to a public audience while identifying and storing no one, and cleared a major studio legal and compliance bar.