Feeding the Machines
We learned to perform for each other online: a clever tweet, a vulnerable confession, a carefully chosen photo. We learned what drew a response and adjusted what we shared. Along the way, public online expression became material that AI developers could learn from too.
The loop is getting harder to untangle. We make things for an audience, platforms decide what to show, and AI helps generate more of what appears in the feed. Our performances become examples for the next round of performances.
Social media is one source of that material, not the whole training library. OpenAI, for example, describes learning from public internet content, licensed material, and information supplied or generated by users, trainers and researchers. That is a broader mix than our feeds alone. OpenAI's account of its training sources.
Still, the social material has a particular character. Threads, captions and comments give developers examples of people writing for a response. We have spent years making ourselves legible to an audience: funny, persuasive, vulnerable, quick to take a position. Some of what AI learns from public writing is how we present ourselves when we know someone is watching.
As more generated content enters circulation, I worry about a cultural feedback loop: familiar styles repeated because they are easy to produce, then used as the starting point for more of the same. There is a related technical risk. Research has found degradation when models repeatedly train on generated data, but the outcome depends on the training setup. Other experiments avoided collapse by retaining original data while accumulating synthetic examples. The evidence supports caution about how that loop is built, rather than a rule that synthetic data always makes models worse. Research on recursive training; research on retaining original data.
The creative question is more immediate: are we adding something worth encountering, or just making the familiar easier to repeat?
For individuals and brands, I would start with three habits.
Differentiate on purpose. Bring a point of view, original observations and constraints that matter to the work. AI can help develop an idea, but you still have to decide what you are trying to say. A fluent result is not a reason to publish it.
Be transparent. Human-made and machine-assisted work already share the same spaces. Be clear about how work was made when that affects the audience's trust. I suspect that knowing a real person stands behind a piece of work will become more valuable as generating convincing material gets easier.
Curate, don't churn. Use the tools to make more considered work, not simply more posts. Choose what belongs in your feed and your brand's voice. The ability to generate another version doesn't create an obligation to share it.
There is also a more intimate version of this problem. People bring private questions, confessions and vulnerabilities to digital assistants. Depending on the service and its settings, those conversations may be used to improve models. There are important exceptions: OpenAI introduced controls to exclude conversations from training in 2023, well before this essay. OpenAI's 2023 announcement.
That makes the old line about being the product if you aren't paying too simple. Payment alone doesn't tell you how your information is used. The service's policies, the kind of account and your choices matter.
What we share, reward and ignore also determines which work reaches other people. Those choices shape our culture whether or not a particular post or conversation ever becomes a training example.
Don't just be a meal for the machine. Be a chef. Decide what is worth making, and take responsibility for what you serve.