Predictive AI Is Hot – But Most Predictions Are Still Flying Blind

01 /The Concept

Artificial Societies has set itself an ambitious mandate: predicting human behavior.

In practical terms, the startup today helps product creators make better product and marketing decisions by running computer simulations of their target audience. The team has been building a suite of tools toward this goal.

Wave came first. The founders used it to create a simulation of 1,000 venture investors and iterated their own pitch deck against that simulated audience – until the pitch was good enough to get them into Y Combinator.

Wave takes a startup description as input and returns two distinct scores: audience favorability and attention. These turn out to be surprisingly different measures that don't always track together.

Beyond the scores, Wave generates a plain-language summary. For Artificial Societies, the verdict looked something like this: "AI-focused investors responded positively, particularly to sections covering growth trajectory and risk mitigation. Investors focused on consumer products were disengaged – the pitch didn't establish a connection to their domain."

Wave also produces specific improvement recommendations. For the same pitch: "Keep using phrases like '100x faster,' 'risk reduction,' 'AI simulation of 1,000 real investors.' Replace 'experimenting with different fundraising strategies' with 'testing different investor outreach approaches.' Concentrate on demonstrating measurable time savings and improvements to startup evaluation quality that the platform delivers to investors."

After getting into Y Combinator, the founders built Reach – a tool that predicts how LinkedIn followers will respond to a post before it's published. Users can refine a draft until the simulated response meets their target thresholds for favorability and engagement.

The two-metric approach turns out to be particularly valuable here. Maximum engagement frequently comes from posts that make people angry. Favorability and engagement pull in opposite directions, so knowing both allows for deliberate positioning.

Artificial Societies claims Reach predicts post performance with 83% accuracy, compared to 17% for a prompt-based prediction from ChatGPT.

A third tool, Form, is in development. It will predict audience reaction to new startup ideas before they're built – with the critical requirement that you define a specific target audience first, since evaluating an idea "in general" tells you almost nothing meaningful.

The platform's database contains over 30,000 AI personas, with more than 5,000 simulations completed at an average audience size of around 1,000 personas each.

Artificial Societies is currently in Y Combinator and published its launch on the YC site yesterday. The startup has received $500K from Y Combinator plus an undisclosed additional amount from other investors.

02 /Why It Matters

On the Artificial Societies site, one phrase stood out: "We analyze the behavior of communities of AI personas" – with "communities" deliberately emphasized. That framing turns out to be the key to understanding why their accuracy numbers beat ChatGPT's so dramatically.

A naive approach to predicting how an audience will respond to a post would model each individual reader in isolation: given this person's known views and interests, how do they react to this text?

The problem is that humans are social animals. Our individual opinions on any given question are heavily shaped by what we see others saying. The same person may hold a very different view on a topic after they've watched a few comments pile up than before any social context existed.

This is one of the most robust findings in social psychology. In a classic experimental setup, subjects were shown two objects – say, one clearly black and one clearly white – then asked to identify the colors after a group of planted participants all gave the obviously wrong answer. A surprising fraction of real subjects, including adults, went along with the group's answer despite direct sensory evidence to the contrary. The expression on the face of the last person to answer – moving from confusion to self-doubt to capitulation – is something to watch.

What Artificial Societies actually simulates isn't a set of individual readers reacting in isolation. It simulates a group reacting in sequence, where early comments influence later ones. The startup built its own audience-segmentation technology, called Hivemind, which clusters audiences by social connection patterns rather than demographic buckets. The current version is labeled 0.1 – the founders consider this the beginning of a long research program, not a finished method.

Even setting aside the social dynamics angle, AI-persona-based audience simulation is emerging as a broadly useful research approach. Polish startup Lakmoos ([related review](/review/mgnovenno-vmesto-polugoda)), which raised €510K, built a platform where product teams can create AI models of their target audience and run instant simulated research sessions – replacing months of traditional user research for product and marketing decisions.

03 /Opportunities

A lot of platforms are emerging around AI-powered prediction. Voyantis ([related review](/review/dolgo-i-mnogo-zarabatyvat-mozhno-tolko-na-teh-kto-dolgo-i-mnogo-platit)) raised $60M to predict customer LTV at the advertising targeting stage. Syrup raised $25.1M for a demand forecasting engine for fashion retail, accurate down to specific colors and sizes. Guac raised $2.8M for grocery demand forecasting focused on perishables. These are just a few examples.

So the broad direction – building AI prediction platforms – is obvious and well-validated. The more interesting question is what makes a prediction platform genuinely competitive rather than one of many.

The Artificial Societies case suggests an answer: the platforms that will stand out are those that incorporate input signals that generic AI systems don't have access to and therefore can't account for. The social dynamics model is exactly that kind of edge – it's a signal source that a ChatGPT prompt simply doesn't capture, and it's the reason the accuracy gap is 83% versus 17%.

For any prediction problem, it's worth asking: what's the input signal that matters most but that a standard model wouldn't know to use? What happens to accuracy if you build in that signal?

Consider the [e-commerce launch platform](/review/zapusk-svoego-biznesa-s-pomoshhju-ii-jeto-uzhe-ne-fantastika) reviewed yesterday. It models expected demand based on product selection, quality, and ad spend. But sales on a marketplace don't happen in a vacuum – they happen in competitive context. Real-world outcomes will be heavily shaped by how many competing sellers enter the same category once they see early traction, and at what price points. A more accurate forecasting engine would model the competitive response, not just the initial market opportunity. And then update recommendations accordingly as the competitive landscape shifts.

In other words, the prediction platform space is more complex and more interesting than it first appears. That's a good thing: it gives platforms built on genuinely better signal sources room to pull away from the pack as the market grows.