A few years ago, putting artificial intelligence into a consumer app was a selling point in itself. Now it's closer to table stakes. Photo editors, messaging apps, shopping platforms, and dating services all run some form of machine learning under the hood, and most users barely register it. The interesting question has quietly shifted. It's no longer whether an app uses AI. It's whether it uses it well.
"Well" is doing a lot of work in that sentence. Responsible AI in a consumer product isn't measured by how much of it there is. A company can pack a dozen models into an app and still treat its users carelessly, and another can use AI thoughtfully and earn far more trust. What separates them is a series of conscious choices, most of which have less to do with the technology than with the decisions made around it. Four of those choices matter more than the rest.
Consent has to come first, especially for sensitive data
The clearest line in responsible AI is consent, and it gets sharper the more sensitive the data becomes. There's a real difference between a model that suggests which photo to lead with and one that analyses your face. The first is a convenience. The second touches biometric data, which is among the most personal information a person has, and which a growing body of privacy law treats accordingly.
The responsible approach is to make anything involving that kind of data genuinely optional, and to be honest about what "optional" means. It should mean the data is never collected, never transmitted, and never processed unless the user has actively chosen to allow it. Not a toggle that's on by default. Not consent pulled out of someone through a confusing prompt. A real, informed choice.
The dating app Agilis offers a useful example. Its photo verification, which confirms that the person in a profile's pictures matches a live selfie, is built so that the facial data is only ever processed when the user themselves chooses to run the check. If they don't opt in, that data is never handled at all. The feature exists to build trust between users, but it doesn't do that by quietly collecting the very thing it's meant to protect.
Where the AI actually runs
The next decision is one most users never see: where the model does its work. An app can send your data to its own servers, or to a third party's AI service, and run the model there. Or it can run the model on your own device, so the data is analysed locally rather than shipped somewhere else to be processed.
This sounds like an engineering detail. It's really a data-protection decision. Running AI on the device keeps the processing local, removes the need to hand personal information to outside providers, and sidesteps a growing thicket of questions about where data physically lives. Governments increasingly expect their citizens' data to stay within their own borders, and every transfer to an external service is another point at which data residency, and trust, can break down. Keeping the work on the device is the cleanest way to avoid most of that.
Agilis takes this approach with the part of its system that decides which profiles a user sees. The model that reads a person's bio and interests to gauge who they might be compatible with runs entirely on their phone. The company's servers perform no AI inference of their own, and the information isn't sent off to a third-party model to be analysed. The AI still does its job, but it does it close to the person it belongs to, rather than in someone else's data centre.
Which model gets used in the first place
There's a quieter choice still, and it's becoming harder to ignore: how big a model an app reaches for. The instinct in a lot of teams is to use the largest, most capable model available, on the assumption that bigger is always better. But large models are expensive to run and consume a great deal of energy, and most everyday tasks in a consumer app don't need anything close to that horsepower.
Choosing a smaller, efficient model that has been trimmed to do one job well is both a practical and an environmental decision. It runs faster and uses less battery, and across millions of interactions it consumes a fraction of the energy a large general-purpose system would. Using AI sensibly means matching the size of the tool to the size of the task, rather than burning resources to look sophisticated. In Agilis's case, the on-device model behind compatibility is a compact one, chosen specifically so it can run efficiently on an ordinary phone without a heavy energy or hardware cost. The restraint is deliberate, and at scale it adds up.
What the AI is ultimately for
If the first three choices are about how AI is used, the last is about why. The most defensible uses of AI in consumer apps tend to be the ones that protect users rather than monetise them.
Safety is the obvious case. Spotting patterns of behaviour associated with scams, such as someone steering a new contact off the platform within minutes or pushing the conversation towards money, is something machine learning does well and humans can't do at scale. Here, AI is working on behalf of the user rather than on them. Agilis uses AI in exactly this way, to classify red-flag behaviour and surface it before it causes harm. An app that points its AI at keeping people safe is making a different bet than one that points it at maximising time on screen or extracting more data from every tap.
Responsible AI is a design philosophy, not a feature
Put together, the through-line isn't any single technique. It's a posture. Responsible AI asks before it touches anything sensitive, keeps data as close to the user as it can, right-sizes the model to the task, and points the technology at protecting people rather than profiling them.
None of this shows up neatly on a marketing page, which is part of why it's undervalued. A consent-first, on-device, right-sized design and a collect-everything, cloud-everything, biggest-model-available design can describe themselves in almost the same words. The difference only becomes visible in the choices a company is willing to make when no one is watching, and in what it's prepared to give up to avoid crossing a line.
As AI becomes the default rather than the exception in consumer software, that's the standard worth holding apps to. Not how clever their models are, but whether they've been honest about what those models touch, careful about where they run, sensible about what they cost the planet, and clear that the person on the other side of the screen is someone to be protected rather than processed.
Jehan Rajendra is the founder of Agilis Dating, a London-based technology company. Agilis is available on iOS and Android worldwide at agilis.dating.
