AI Training Data Shifts Toward Paid Human-Sourced Data
Summary
YouGov (YUGVF) and Zedge (ZDGE) say AI labs and robotics developers now pay for human survey and task-video data that cannot be scraped, moving margin to panel owners.
AI labs and robotics developers are becoming paying buyers of human data that cannot be scraped from the web. Between October 13 and October 26, 2026, online research firm YouGov plc (YUGVF) and app company Zedge, Inc. (ZDGE) both said on their earnings calls that this new class of customer is commissioning survey answers and task video that only a network of real people can produce.
Why AI developers have to buy human data that cannot be scraped
Much of the data used to train earlier AI models was collected by scraping public web pages. Two kinds of demand now have no ready-made supply online. Frontier labs such as Anthropic and OpenAI run very large surveys to understand their users and the public [1]. Robotics developers need first-person (egocentric) video of people doing tasks, recorded with synchronized sensor data, so their models can learn how people handle real physical environments [2].
Both kinds of data have to be produced by real people to a specification, and AI companies find it hard to organize those people themselves. YouGov runs a long-standing online panel of respondents who agree to take part in research. Zedge, through its DataSeeds.ai unit, organizes contributors who shoot material to a customer's technical brief. If more AI developers run into the same gap, this kind of order could flow to other companies that own a trusted panel or a capture crowd.
YouGov and Zedge report new AI clients and larger orders
On its FY26 results call (year to July 2026), YouGov management described AI as a new vertical and said Anthropic and OpenAI have become big clients. Dedicated AI companies contributed about GBP4 million of revenue; including other clients' work that was triggered by AI, related sales over the last 12 months came to about 11 million (the transcript renders the currency incorrectly; from context it is almost certainly GBP) [1]. Management said these clients do not do the research themselves, care little for advice or packaging and want the data raw, and that YouGov's margin comes from the quantity of high-quality data [1].
On its FY26 fourth-quarter call, Zedge said it has started delivering data in robotics, that order sizes are increasing, and that a couple of prospects in its pipeline are discussing six-digit-plus orders [2]. Management added that these prospects want dedicated data sets they can own outright, without the data reaching their competitors [2].
One company sells survey responses and the other sells custom video, yet both show AI developers paying directly for freshly collected human data. Both figures have limits. YouGov's GBP4 million is traditional reputational research that the labs use to understand audiences, not model-training input [1]. Zedge does not report DataSeeds revenue separately [2].
Margin shifts toward whoever owns the human network
As both companies describe it, AI buyers rarely pay for advice or packaging; what they pay for is raw data that only the owner of a panel or a contributor crowd can supply [1][2]. For this demand, consultancy-led custom research and static image-library licensing are less likely to capture the spend, while companies that hold the direct relationship with respondents or contributors gain bargaining power.
The opposite force is also visible. YouGov itself launched Parallels, a product that supplements human data with synthetic "digital twins" [1], and Salesforce (CRM) is acquiring AI customer-research platform Listen Labs for $2 billion [3], so synthetic respondents could absorb part of the survey demand. Ipsos (IPSOF), which also owns panels, disclosed no AI-lab revenue on its Q3 2026 call and reported organic growth of only 2% [4]. YouGov's overall revenue growth was only low single digit [1]. Signals to watch are whether YouGov's AI-related revenue keeps growing and whether Zedge's six-digit-plus orders close.
Companies exposed to the same human-data demand
- Innodata (INOD): The data-services company collects and annotates data for AI clients; in Q2 2026 its egocentric video-collection pilots with robotics companies moved toward a roughly 2 million hour program that has not yet been awarded, placing it in the same commissioned-collection step [5].
- TaskUs (TASK): The outsourcing provider's AI services revenue was $66.1 million in Q2 2026, up 26% year over year, and it uses its Taskforce platform to deploy specialized crowds that collect egocentric data, the same kind of collection Zedge describes [6].
Sources
[1] Drillr · YouGov plc · 2026-10-13 · FY26 results call
[2] Drillr · Zedge, Inc. · 2026-10-26 · FY26 Q4 earnings call
Much of that data cannot be scraped from the internet or acquired from an existing library. It must be produced subject to a detailed technical brief.
[3] Drillr · Salesforce · 2026-09-09 · News
[4] Drillr · Ipsos · 2026-10-23 · Q3 2026 earnings call
[5] Drillr · Innodata · 2026-08-06 · Q2 2026 earnings call
[6] Drillr · TaskUs · 2026-08-05 · Q2 2026 earnings call
This analysis highlights industry changes and companies that may be overlooked. It is not a stock recommendation.