Yair Adato

The through-line

Adato has been repeating the same two words for five years: responsible and open. "I started five years ago and I said that look guys generative AI is going to come in and it must be responsible and must be open" [1]. Responsibility, in his definition, is not a values statement but a design constraint: "responsible it means that you take accountability of the technology and the product that you build" [1], which in practice means fully licensed training data, no scraping, no deepfakes, no celebrity impersonation, and attention to bias and fairness [1]. Openness is the commercial counterpart: if the technology stays in the hands of a few labs, "if only open Ai and Google have access to this technology there will be very very non-innovative world" [1].

What has moved over time is where he locates the difficulty. Early on the emphasis is on what generative AI makes possible for people who need visuals and cannot get them [2]. Later the emphasis is on why possibility is not adoption: the model quality has run ahead of everything else, so the binding constraints are legal compliance, pricing structures, business model change and trust [1]. "Technology is only one aspect of it and in order to have something really adopted by by Enterprise you need to see the full picture" [1].

On what responsible AI actually means

Adato refuses to let responsibility float free as a slogan. He grounds it in specific, checkable properties of the product: "we build a platform that respect copyright and privacy every single visual that we generate uh we have uh all the training that we use in the training set we have a full license for that so we didn't scrap the internet we didn't steal data" [1]. Deepfakes and celebrity generation are blocked [1]. He treats copyright and privacy as the entry point rather than the whole of it: "there there are Ms bias and fairness and other element of that but to start with respect copyright and privacy regulation" [1].

He is unusually blunt about rhetorical tactics he sees in the industry. Some strong voices, he says, try to convince you that legal compliance and responsibility are not important, and others shift the ground: "they will start the discussion about responsible Ai and before you notice they're talking about safety no safety is another topic don't don't don't confus use the audan" [1]. He points to a CEO who claimed there was not enough evidence that copyright matters to users, and reads the motive as commercial [1]. His own instinct is personal as much as strategic, framed around his fifteen-year-old twin daughters and TikTok: "I don't want to be the one that create the next Tik Tok" [1].

On bias he takes the same engineering line rather than a moral one. The first step is acknowledging the problem exists, which nobody did five or seven years ago [2]. After that it becomes routine: "it's like security you don't solve for security it's not like okay today we are going to do security you understand that there are some standard that you need to meet" [2]. You check that the training and evaluation sets are unbiased, and you compare results across subgroups, and it becomes inherently part of the system, like privacy [2].

On attribution and the licensed data economy

The mechanism behind the licensing claim is what he calls a Nano attribution engine: "every time that we generate an image we know how to trace back in the training set which are the visual that impact the most for this generation and we pay royalty for them" [1]. Revenue is split back to the data owners on behalf of the client, which is what allows back-to-back agreements with the major stock image providers [1]. A side benefit he stresses is quality, since stock repositories are built for exactly the marketing use case the output is destined for [1].

He reads the current wave of lawsuits, the Scarlett Johansson dispute with OpenAI, and the letters sent to music startups as a phase rather than an endpoint, and the analogy he reaches for is music streaming: it began with companies scraping all the music in the world, then "people said it's not cool it's not cool to steal from the artist and and not pay for them and then the economy involved and then come new model for example Spotify" [1]. The conclusion he draws is commercial, not moral: "you must have a solution in order to have a sustainable product to sustainable economy" [1]. He is alert to the practical exposure this creates for buyers, citing a large retailer that ran a major AI project and only discovered afterwards that Johansson had accused OpenAI of taking her voice [1], and a Figma design feature that produced output resembling Apple's OS closely enough that it had to be pulled [1].

On whether machines can be inspired the way artists are

Pressed with the argument that a musician absorbing every blues record is itself a neural network, and that generative models are only doing what humans have always done, Adato gives two answers. The first is transactional and, in his view, decisive: the musician "buy this music he go to a concert he buy the ticket he buy the album he buy to Spotify" [1]. The second is about output quality. He grants that some philosophers hold that we are just very advanced algorithms, and nods to Harari on the data religion, but insists that the musicians he has met "have soul when they create the music" [1], while what current systems produce "is flat it's it's average it's it's a machine that can create uh boring uh uh boring content" [1]. His forecast follows: "we are going to be in a world of boring content" [1]. None of this makes him hostile to the tool. It is a good tool, people will use it, and it will help people make more music and more content [1].

On why the frontier is not where the value is

His most quotable position is that "generative AI is reached its peak and it's now a race to the bottom" [1], and he means it narrowly: out-of-the-box text-to-text and text-to-image have no value left in them. He calls this the horses-on-the-moon scenario, amazing to look at and useless to him: "it's not what I need" because it does not contain his brand [1]. Value sits in the opposite properties. "In order to create something with value you need to have something which is very controllable predictable follow your brand follow your style" [1], with e-commerce as the limit case where you cannot alter even half a tone on the product [1]. He also warns against the reverse error of technology-led product: "sometimes we eat the technology no there a great technology use that that's okay" [1].

The same instinct shapes how he describes the underlying need in visual work. Stock libraries do not fail because images are missing but because they are nearly right: "it's never the image that they need it's almost always almost the image that they need" [2]. The colour is wrong, the facial expression is wrong, the presenter is wrong, the background needs to be more urban. "It's always 80% what I need" and then it goes to a designer who is slow, and who sometimes cannot change the presenter or the background at all [2]. Illustration behaves differently, since a raccoon in space eating sushi has to be generated rather than found, but even there the first sketch is the beginning of the negotiation, and the shirt still has to match brand colours [2]. His summary of the product is therefore not creation but adjustment: "it's not about creating new visual it's about to adjust the visual to the goal of the communicator" [2].

He defines the user accordingly. The communicator is "someone with a business goal which need visual in order to accomplish this visual goal" [2], spanning marketing, presentations, internal and external communication, sales collateral and e-commerce [2]. The promise is that anyone can generate or modify visuals for a business goal "only by AI without a camera and without a designer" [2], while designers themselves get to make things that would be much harder otherwise, because the AI removes a technical barrier for both [2].

On selling to enterprises, and why it is slow

Legal comfort is the door, not the deal. "Having the legal compliant the legal department happy is the is the first step then there is the second step of okay so now I can use it what does it mean for my business what does it mean for my business model" [1]. An ad agency that used to sell one thing now has to sell something slightly different, and organisations find that kind of change hard [1]. Then comes pricing, where he is scathing: the models on offer are "extremely naive like extremely extremely naive" [1]. Consumption pricing is inherited from Nvidia, AWS and Microsoft, all partners he praises, but it does not fit the buyer: "I don't want to to pay every time that I click a button of changing something in my web tool this is not how I'm as a consumer pay" [1]. He argues for the flip to flat fee or per seat, and reads the pricing confusion as a symptom of a deeper unresolved question about where the real value of generative AI actually sits [1].

He also sorts prospects by readiness. Some executives arrive at the meeting saying they must do something with AI, and the test he applies is simple: send the presentation of the use cases you want to solve. If it comes, the conversation continues. If not, "we will call them three months from today they're not ready right" [1]. And he is patient about timelines because he has been wrong before in the same direction: at his first computer science conference in 2007 the opening lecture was on autonomous driving, and he was certain his daughters would never need a licence. They are fifteen and they are going to get one [1]. "It's not that easy it will take time companies need to invest resources to create amazing product" [1].

On being the picks and shovels

The go-to-market follows from the openness argument. The web tool exists but is not the point: "the main goal is to use use it as an API or as a source developer we give companies the ability we give the code so they can integrate that into their processes and product" [1]. He describes large organisations with real engineering and AI teams building their own promotions, ads, marketing material and augmented reality experiences on top of it, and gaming companies giving their own users tools to generate 3D assets, scenes and worlds [1]. "We are the ENA we are the Pix and Shel for this technology" [1]. He frames the expansion beyond stills the same way, with video, music, audio and 3D planned within the following two or three quarters, on the principle that every type of asset comes with rights attached [1], and he has flagged premium features and integration with a major stock provider as near-term releases [2].

The consumer market he deliberately declines. His generation targets photorealism, so the output looks like a photograph of a real person, and handing that to consumers invites customisation "in a way which not reflect the reality" [2]. His concrete fear is a twelve-year-old modifying a photo of a friend and using it for shaming [2]. Beauty filters fall into the same bucket. "I aim for corporate businesses" [2].

On trust as the real adoption problem

Because the technology reads as black magic, Adato thinks the scarce resource is trust rather than capability. "People need to build trust with the system they need to trust the system it's super important there are so many small tricks that all of us doing to create that it will be trustworthy and once you break this trust it's really hard to recover" [1]. The user's questions are self-interested and legitimate: does it work for me or for someone else, will it leave me without a job [1]. He extends this to the public at large, arguing that after twenty years of AI people will no longer take what engineers and companies say at face value, and that they are right to ask how it works, why it was built this way, how the data was gathered, whether the results are unbiased, and whether the system improves society [2]. Aggression on the vendor's side is counterproductive: "when you are too aggressive it's a problem" [1]. His own posture toward the technology is consistent with this. Fear of AI is understandable, and the answer is not to be afraid but "to mitigate it to manage it to have a responsible AI when we build product" [2]. He applies the same conscious judgement to himself as a user, refusing TikTok for reasons he investigated rather than reasons of ignorance, and treating the virtual world exactly like the physical one, where he does not go everywhere on the planet either [2].

On automation, jobs and where this revolution sits

He rejects the framing that hype and significance are alternatives. "It's a hype but it's out doubtfully shift the human behavior it's a huge Revolution the fact that it's hype doesn't mean anything" [1]. His historical frame is that all technology is automation, and that the previous two waves automated the copying of data, through printing and CDs, and then the distribution of data, through the internet and mobility. This one automates the generation of data, and he thinks it is the stronger of the three [1]. The limit of foresight is his real point: in 1997 nobody imagined remote interviews, e-commerce or Airbnb, because "we were limited in the way we thought about data" [1], and the same failure of imagination applies now.

On employment he is unsentimental and steady across both conversations. Technology always automates a process and people adapt to it, as they did when cars displaced horses and created drivers, and as they will when autonomous vehicles displace drivers [2]. "People will not lose their job they will need to adjusts to the new technology" [2], with the qualification that those who stay behind and do not adapt will have a problem, "like always there's nothing new" [2]. He is equally quick to deflate the notion that the machine is about to think for us: it is far more complicated than people imagine, and that future is not here yet [2].

On building an AI company

His first piece of advice to founders is not technical. "Spend as much as time as you can to understand the problem that you want to solve and the market that you want to to serve" [2], because changing markets later is possible but hard and changing the problem is harder still. Work out who has the problem, why, how much they will pay, and what alternatives they already have, and "don't rush yourself" [2].

He insists that AI development is a different discipline from engineering, and that conflating the two is the standard failure mode of founders with a good idea who do not understand the technology. Nobody is surprised that hardware and software are built differently, and he wants the same acknowledgement for AI: the architecture questions differ, the things you build differ, and "the expected output from AI researcher after three month three weeks is different than what you expect from an engineer" [2]. His prescription is to bring in AI talent as early as possible and give them the freedom to work [2]. The challenges cluster around research immaturity, data access and talent, plus the frequently underrated task of turning working AI into a product [2].

He is dismissive of AI as a pitch ornament, listing it alongside blockchain, virtual reality and bitcoin as buzzwords [2]. Many companies do not need it: "too expensive too complicated too risky" [2], and a rule-based decision layer is often exactly what the situation calls for, especially where the company is not a pure AI play [2]. He singles out the data-hoarding pitch for particular scorn, the claim in due diligence that a company will collect data and monetise it later, "which means that we have a lot of data we have have no clue what to do with it and we hope that someone will pay for us" [2]. For those who genuinely need AI, he lists the infrastructure questions to settle in advance: how you will train, how to stop training from consuming the seed money, how you will measure results, how to balance a fast signal that you are heading the right way against a slower and riskier path to production quality, and what your failure strategy is, because "it will not be 100% working in the first year" [2]. Underlying all of it he names one technical challenge as central to generative AI specifically: how to measure the quality of the output automatically or semi-automatically [2]. He also flags cross-platform generalisation as a persistent difficulty, since systems that work on one task often do not transfer [2].

Takeaways

• Responsibility is defined operationally: fully licensed training data with no scraping, blocked deepfakes and celebrity generation, and bias treated as a standing requirement like security or privacy rather than a one-off project [1][2].
• The Nano attribution engine traces which training visuals most influenced a given generation and pays royalties back, which is what makes back-to-back licensing with stock providers possible [1].
• The music streaming arc is his template for where this ends: scraping, backlash, then a licensed model like Spotify, and companies without a licensing answer will not have a sustainable business [1].
• Out-of-the-box text-to-image is a race to the bottom, and the remaining value is in output that is controllable, predictable and on-brand, which is why he sells to enterprises through APIs and code rather than as a consumer toy [1].
• Legal comfort only opens the door; the harder blockers are business model change and pricing, where consumption-based models inherited from cloud vendors are, in his view, extremely naive and should flip to flat fee or per seat [1].
• He declines the consumer market on purpose, because photorealistic editing in the hands of a twelve-year-old invites shaming and abuse [2].
• Trust is the scarce resource, easy to break and very hard to recover, and users are right to interrogate how a system was built and whose interests it serves [1][2].
• For founders: understand the problem and the market before building, hire real AI talent early because AI development is not engineering, and be honest that many products need a rule-based system instead [2].