OpenAI admits AI text can become undetectable after ordinary edits, while Ursula von der Leyen’s EU rules are already live — real need, ugly startup shape
This week’s pick: Best bet if you have money and an audience
What you’re reading: a real Opportunity Dossier from our AI research pipeline — Ada spotted the signals, Bea shaped the candidates, Cora dug deep, Edie tried to kill it and owns the score, and Della wrote it up. New to our scoring? Start with How Our Star Score Works. Research date: October 2026 — facts move; check dates before betting money.
OpenAI says watermarking is coming for eligible European Union users. The European Commission says the transparency rules are already applicable. Adobe says people need a way to tell what is behind content. Hany Farid, the best-known public deepfake skeptic in the field, keeps pushing the market toward provenance instead of gut-feel detection. The fight is not whether the problem exists. The fight is whether a startup can own enough of the solution before Adobe, Microsoft, Google, Meta, and OpenAI swallow the category whole. OpenAI
“Over the coming weeks, we will introduce text watermarking to eligible ChatGPT and Codex users across all plans in the EU only.” OpenAI
“Text generated with OpenAI tools may be too short, edited, or translated for detection to work reliably.” OpenAI
“Europeans have a right to know whether what they see, hear or read has been made or altered by AI, especially when such content can shape public debate. Transparency is how we protect trust.” Ursula von der Leyen said that when the Commission published the Code of Practice on marking and labelling AI-generated content. European Commission
“At this critical moment in history, as generative AI becomes more powerful and prevalent than ever, people need a way to tell what’s behind the content they’re consuming.” Dana Rao, Adobe’s General Counsel and Chief Trust Officer, said that when Adobe was pressing the case for Content Credentials. Adobe
A stranger should care because a boring compliance rule just collided with a messy technical truth: the labels are now required, but the evidence often breaks under real-world use. An investor should care because the pain is real but the capture is narrow. A founder should care because the obvious version of the business — “we do watermarking for EU compliance” — is a trap. The only plausible company here is a cross-vendor workflow and audit layer.
The law is real, but the technology is openly brittle.
The giants already ship most of the ingredients, so the wedge is much narrower than the pitch sounds.
The survivable company is an evidence-and-workflow control plane, not a generic detector.
Provenance and watermarking services for EU compliance = software that helps a company mark, detect, label, verify, and log AI-generated content well enough to satisfy internal policy and external regulators.
The Allure
The believers’ dream is easy to understand. Europe moved from talking about AI transparency to enforcing AI transparency. The Commission says Article 50 transparency obligations, including marking and labelling AI-generated content, are applicable from 2 August 2026. A rule with a date tends to wake up budget holders in a way white papers never do. European Commission
The dream gets stronger because the harm is no longer abstract. Nieman Lab reported in October 2026 that fake New Yorker-style cartoons generated with ChatGPT carried the signatures or pen names of real cartoonists, including Brendan Loper, and one related tweet drew about 25,000 likes. A rights problem, a reputation problem, and a trust problem all landed in one ugly little media incident. Nieman Lab
The investor story, at first glance, sounds attractive. Compliance budgets exist. Trust-and-safety budgets exist. Brand-risk budgets exist. Newsroom-integrity budgets exist. If a startup could become the place where enterprise content teams check incoming assets, apply publish rules, and store evidence for later disputes, the startup could sell annual contracts instead of fighting for consumer subscriptions. The likely first customers are publishers, enterprise creative-operations teams, digital asset management teams, content management system owners, agency compliance groups, and platform trust teams. Adobe Business
The believers also have a real technical story. Provenance is stronger than pure “spot the fake” detection because provenance starts at creation. A content file can carry machine-readable information about how the content was made or edited. The Coalition for Content Provenance and Authenticity, usually shortened to C2PA, exists to standardize that record. Microsoft announced in 2021 that Adobe, Arm, BBC, Intel, Microsoft, and Truepic were joining forces to create the C2PA standards group. Microsoft
Then comes the part that makes founders lean forward. Every big vendor is building only part of the stack, and each stack has limits. OpenAI is rolling out EU text watermarking, but OpenAI also says edited, translated, or short text may defeat detection. Microsoft says provenance detection verifies Microsoft-issued signals only and should not be treated as proof of authorship, ownership, or endorsement. Meta can read outside markers for image labels, but only after those markers exist and survive the trip. Google has a verification portal, but Google’s portal checks for Google’s watermarking system and partner coverage, not a magical universal truth machine. The gap between all those partial systems is where a startup starts daydreaming. Microsoft Learn
If the dream works, the company does not win by inventing a better invisible mark. The company wins by becoming the boring adult in the room: the software that knows which evidence to trust, when to label, when to escalate, what to retain, and what to show a regulator six months later. That kind of software can become sticky inside publishing and moderation workflows. That is the dream.
Where This Idea Came From
The roots of the idea run through standards people, not startup people. Adobe lit much of the early fire with the Content Authenticity Initiative in 2019, then helped co-found C2PA in 2021 with Microsoft and other partners. Adobe kept selling the vision that content should “tell its story” through attached provenance records rather than leaving viewers to guess. Adobe Blog
Microsoft joined that standards push early, but Microsoft’s current product material is more cautious than startup decks usually are. Microsoft says provenance signals can show that Microsoft systems generated or processed content, while also warning that provenance signals do not prove ownership, authorship, intent, or endorsement. Microsoft is not promising magic, which is useful honesty and terrible news for any founder tempted to overclaim. Microsoft Learn
The newer accelerant came from regulation. The Commission’s Code of Practice on transparency for AI-generated content turned provenance from a standards conversation into an operations deadline. Ursula von der Leyen put the political framing in plain language: Europeans have a right to know when content was made or altered by AI. That sentence matters because compliance buyers hear “right to know” and imagine future scrutiny, not just present paperwork. European Commission
OpenAI then poured fuel on the fire by doing two things at once. OpenAI announced EU-only text watermarking for eligible ChatGPT and Codex users. In the same post, OpenAI admitted that edited, translated, or short text may be undetectable and said detector access would be granted on a case-by-case basis. OpenAI essentially described the startup opportunity and the startup fragility in one document. OpenAI is shipping a compliance-facing mechanism while warning buyers not to trust the mechanism too much. That is a real self-tension, and the self-tension is the whole market. OpenAI
The human fight underneath the category is now obvious. Regulators want labels. Vendors want their own standards and tooling to win. Publishers and enterprise content teams want an answer that survives audits. Creators want attribution that does not falsely accuse human work of being machine-made. Journalists and rights-holders want fewer Brendan Loper moments. Nobody actually disagrees that trust is breaking. The disagreement is about who controls the repair.
Who Wants to Build It
The surprising part is how thin the obvious founder swarm looks. A hot category usually produces a flood of accelerator demos, builder threads, and “we’re the Stripe for X” launches. Provenance-for-compliance has some of that, but not much. The visible energy is stronger inside incumbents and standards bodies than inside a loud startup wave. C2PA
Some adjacent startups are clearly in the room. Truepic raised a $26 million Series B in 2021 led by M12, Microsoft’s venture fund, around secure camera and authenticity workflows. Truepic is more capture-chain integrity than “EU compliance middleware,” but Truepic proves that buyers will pay for trust tooling in regulated workflows. Truepic
Reality Defender raised a $15 million Series A in 2023 to detect deepfakes across text, video, image, and audio. Reality Defender is adjacent rather than identical, but adjacency matters because many buyers may frame the budget line as “deepfake detection” instead of “provenance compliance.” A startup selling into the same trust-and-safety or security buyer must expect adjacent vendors to show up in the deal. TechCrunch
GetReal, co-founded by Hany Farid, raised $17.5 million by March 2025 and positioned around deepfake detection and authentication. Hany Farid matters less as a direct quote machine for this startup and more as a public signal that serious operators see trust in media as a real budget area. TechCrunch
Our read: founder demand exists, but founder demand is thin and careful. Thin founder demand can mean hidden opportunity. Thin founder demand can also mean everybody smart already noticed the sales cycles, standards politics, and feature risk.
Scorecard
Time-to-Truth — 2/5
Time-to-truth means how fast a founder can get a trustworthy buy-or-don’t-build answer. The cheap version takes 6 to 12 weeks and about $3,000 to $15,000 through 20 to 40 buyer interviews plus 5 to 10 design-partner conversations with named humans such as heads of content operations at publishers, digital asset management owners, content management system leads, and trust-and-safety managers. The honest version takes longer because one-page interest is not enough. Real truth appears only after 1 to 3 pilot integrations into enterprise workflows, where false positives, evidence gaps, and audit requests show up.
Time-to-MVP — 2/5
Time-to-MVP means how long it takes to build the first usable product. A narrow first version takes about 3 to 6 months with 2 to 4 technical people if the wedge is cross-vendor verification, policy routing, and evidence logs rather than brand-new watermark science. An enterprise-ready version with security controls, data retention, and workflow integrations is more like 6 to 9 months.
Time-to-First-Customer — 1/5
Time-to-first-customer means how long a cold stranger waits before a meaningful paying customer appears. Expect 6 to 18 months for a cold founder. A founder with warm relationships may compress the cycle to 3 to 6 months, but the default case is slow because the likely buyers are cautious enterprise teams with review cycles, legal review, and integration concerns.
Capital Requirements — 2/5
Capital requirements means both what the founder personally puts in and what the business really needs. Founder cash can start around $15,000 to $75,000 for a narrow pilot product, travel, basic security work, and founder-led selling. A serious first year aimed at EU compliance buyers likely needs $250,000 to $1,200,000 total for engineering, legal and privacy work, storage and processing, security, and consultative sales (estimate).
Execution Complexity / Operational Risk — 1/5
Execution complexity means how many ugly operational problems sit between the pitch and reality. This category is loaded with them: shifting vendor application-programming interfaces, meaning the interfaces software uses to talk to another company’s systems; uneven detector access; edge cases across media types; metadata stripping; ambiguous policy decisions; and enterprise security requirements. Microsoft’s own material warns that provenance detection does not determine authenticity and only verifies Microsoft-issued signals, which limits what any dependent startup can promise. Microsoft Learn
Competitive Intensity — 1/5
Competitive intensity means how crowded and dangerous the field already is. Adobe, Microsoft, Google, Meta, and OpenAI all ship provenance, watermarking, labeling, or verification components. Adjacent venture-backed firms already chase similar buyers. A startup is not entering an open field. A startup is trying to squeeze into the cracks between much larger stacks.
Market Momentum — 4/5
Market momentum means whether demand has a live forcing function right now. The answer is yes. The transparency obligations are applicable from 2 August 2026, and the Commission has already moved from drafting to rollout. That is real momentum, not theory. European Commission
AI Leverage — 3/5
AI leverage means whether AI meaningfully reduces cost or improves the product. AI helps with fallback detection, triage of ambiguous cases, evidence summaries, and policy suggestions. AI does not create the moat. The real value is in workflow placement, orchestration, and retained evidence.
Defensibility — 2/5
Defensibility means how hard the business is to copy or absorb. A generic watermarking or detection product is weakly defensible because incumbents can add the feature. Some defensibility can form if the company becomes the system of record for provenance evidence across vendors and sits deeply inside content approval, moderation, or publishing workflows. That moat forms slowly and only with buyer trust.
Key Metrics That Matter
- Founder personal cash: $15,000 to $75,000 for an initial wedge, pilot support, travel, and basic security work (estimate).
- Total capital needed: $250,000 to $1,200,000 to build a credible enterprise business (estimate).
- Time to break-even: 12 to 30 months to cover costs, assuming 3 to 8 enterprise customers paying roughly mid-four to low-five figures annually and service overhead stays controlled (speculative).
- Audience you’d need first: about 150 reachable decision-makers or strong influencers in European Union-facing content operations, publishing, trust-and-safety, digital asset management, content management systems, and regulated brand workflows. “Audience” here means people who already know and trust the founder enough to take a meeting, not strangers from ads.
- Audience basis: 150 reachable buyers × 10% serious evaluations × 20% close rate × $15,000 to $40,000 first-year contract value = about $45,000 to $120,000 initial annual recurring revenue cohort. One cohort does not solve the business. Several cohorts or larger contracts do.
Your Money
The founder personally invests $15,000 to $75,000 at the start. The business likely needs $250,000 to $1,200,000 total before the company looks credible to enterprise buyers (estimate). That gap matters. A founder can prototype without a venture round. A founder probably cannot build the whole company shape without real capital or unusually patient customer funding.
Time to Break-Even
Break-even likely takes 12 to 30 months (speculative), assuming the product lands 3 to 8 enterprise customers paying roughly $15,000 to $40,000 per year and consulting overhead does not balloon. The load-bearing assumption is that buyers treat the software as compliance workflow infrastructure, not as a nice-to-have detection gadget.
The Audience You’d Need First
The founder would want roughly 150 warm, reachable buyers or buyer-adjacent humans inside publishers, enterprise creative-operations teams, trust-and-safety teams, digital asset management teams, content management system teams, and regulated brand-content operations. Without that audience, add roughly 6 to 12 months of relationship-building before pipeline becomes real (estimate).
What Moves the Needle
For a cold stranger, the business is 2 stars at best and often behaves like a 1-star practical choice because the time drain is so high. For a founder who already has the audience, the rating rises to 3 stars because warm access can get pilots before incumbents close more of the gap. With more money alone, the rating stays 2 stars because capital cannot remove feature risk against Adobe, Microsoft, Google, Meta, and OpenAI. With both money and audience, the ceiling is 3 stars. No amount of money or audience fixes the core cap: native vendor tooling can absorb much of the standalone value for generic watermarking.
Who Should Build This — And Who Shouldn’t
The person who wins here is a patient founder who has already lived inside enterprise content workflows, trust-and-safety, digital asset management, newsroom systems, or compliance-heavy software — a founder with enough technical fluency to avoid fake promises and enough relationship capital to get cautious buyers into pilots.
The person who’ll be tempted but shouldn’t be is the talented solo builder who sees a fresh European Union rule and assumes regulation automatically creates fast software demand, or the creator-tool founder who wants to sell a clever verifier interface while dodging procurement, integrations, and policy ambiguity.
What It Would Take to Raise This Score
A founder would need to prove a narrow wedge that incumbents are not serving: cross-vendor evidence orchestration inside a messy workflow. The best score-lifting evidence would be 2 to 3 paid pilots at $5,000 to $20,000 each with publishers, trust-and-safety teams, or digital asset management owners who already use multiple AI vendors and cannot manually stitch native tools fast enough. The founder would also need signs that buyers want retained logs, policy routing, and audit evidence — not just another detector dashboard. If repeated interviews show that most prospects either stay in one vendor stack or are happy to stitch native tools manually, the score does not rise.
Go-to-Market
The best channels are not glamorous.
1. Cold outreach and direct sales — strong fit. The buyer universe is small and nameable, so the founder should target specific publishers, enterprise creative-operations leaders, digital asset management owners, content management system leaders, and trust-and-safety heads. Realistic early cost is mostly founder time plus modest tooling, about $300 to $1,500 per qualified conversation and $5,000 to $25,000 in founder travel, demo work, and pilot support to land a first customer.
2. Partnerships and plugin directories — strong fit. The best compounding path is to live where provenance decisions already happen: Adobe ecosystems, content management system ecosystems, digital asset management ecosystems, moderation stacks, and agency compliance workflows. Integration pages and partner directories are not sexy, but embedded placement is the closest thing this business has to leverage.
3. Associations, trade shows, and niche events — strong fit. Trust-and-safety events, newsroom-integrity gatherings, digital asset management conferences, Adobe and creative-operations events, and relevant trade associations concentrate exactly the cautious humans who feel the pain and ask peers what is acceptable.
First 90 days should look like this: in weeks 1 to 3, build a target-account list of 50 to 100 named organizations and create a “provenance gap audit” offer; in weeks 2 to 6, run 25 to 40 interviews and publish a short field report on where native tooling breaks in multi-vendor workflows; in weeks 4 to 8, ship one narrow prototype that checks signals and logs decisions; in weeks 6 to 12, try to close 2 to 3 design partners for paid or clearly conversion-scoped pilots.
One thing a day for 90 days: send one targeted outreach note to a named buyer, comment on one regulation or provenance update, request one buyer conversation, publish one teardown of a real workflow gap, or advance one partner conversation.
Considered and set aside: search content, AI-answer-engine authority, paid search, paid social, organic social as a primary channel, influencer, email-list-first, own podcast, community, guest podcasting, newsletter swaps, affiliates, press-first, Product Hunt, awards-first, app stores, listing sites, direct-message-only selling, local press, freemium, virality, sampling as the main wedge, and referral programs.
Illustrative Case Study — not a real company
A fictional founder named Elena spent six years inside a European publisher’s content systems team. Elena knew the actual pain: one desk used Adobe workflows, another desk used OpenAI tools for drafts, freelance illustrators used a mix of third-party image tools, and social teams stripped metadata without noticing. When legal asked, “Can we prove what was AI-generated and what was edited by staff?” nobody had a clean answer.
Elena did not start by promising universal detection. Elena offered a paid 6-week provenance gap audit for one publishing pipeline at $8,000. The audit found three concrete failures: one image workflow dropped attached credentials during resizing, one text workflow depended on OpenAI output that became unreliable after translation, and one moderation queue had no retained evidence for why a label had been applied.
The pilot software did only four things: collect signals from supported vendors, flag unsupported or stripped assets, apply simple rules such as “label,” “escalate,” or “hold,” and store an evidence log. The pilot succeeded only because Elena already knew the humans who owned publishing operations and legal review. A stranger founder with the same code but no warm access would still be trying to get the first meeting.
That illustrative scenario shows the narrow truth: the product is not “spot the AI.” The product is “keep the messy workflow from embarrassing the company later.”
The Honest Skeptic’s Case
The skeptic’s case is brutal and mostly fair.
First, the category can collapse into features. Adobe already markets enterprise content-authenticity products based on Content Credentials and C2PA. Microsoft has provenance and disclosure tooling. Google operates SynthID Detector. Meta labels outside-party AI images by detecting signals from multiple vendors. OpenAI has a verification tool and a content provenance interface for image and audio, plus restricted text detection for European Union use cases. A generic startup pitch that ignores those facts is fantasy. Adobe Business
Second, the core technical promise is weaker than a founder wants it to be. OpenAI says ordinary edits, translations, or short passages can make AI-generated text hard to detect. Microsoft says provenance does not establish authorship, ownership, intent, or endorsement. That means a startup inherits the limits of upstream tools while still taking blame when a buyer expects certainty. OpenAI
Third, the buyer may not exist in one neat budget bucket. Is the buyer a compliance officer, a newsroom systems head, a trust-and-safety lead, a chief information security officer, a brand-risk manager, or a digital asset management owner? The answer changes by account. Category ambiguity slows sales even when pain is real.
Fourth, our research could not honestly verify a clean total market size for “AI provenance compliance services.” That gap matters. “Big problem” is not the same as “big software category.” We also do not have clean, sourced average contract values from direct peers in this exact slice, so any financial model here carries real uncertainty.
The Thesis-Killer Alternative
The thesis-killer alternative is simple: use the native tools.
A buyer can stay mostly inside one vendor stack, rely on Adobe or Microsoft or OpenAI or Google tooling, use platform labels where available, and handle the rest manually. For many organizations, that path may be good enough and vastly cheaper than buying a new control layer. That bypass absolutely kills the broad startup thesis.
The broad startup thesis fails because the giants already provide generation-time provenance, verification portals, or platform labels. If the customer’s workflow is simple and mostly single-vendor, the startup is unnecessary. Meta
The narrower thesis survives only where the buyer lives in a mixed environment and actually cares about retained evidence. OpenAI’s limited detector access and admitted text fragility help the narrow thesis. Metadata stripping and normal distribution changes help the narrow thesis. The need to explain missing or ambiguous evidence during a later dispute helps the narrow thesis. So the bypass does not defeat the wedge of cross-vendor compliance workflow plus evidence retention. The bypass does defeat almost everything wider than that.
Business Model, MVP, Time-to-Truth Test, Risks, Reasons Not to Pursue
The business model should be annual business-to-business software contracts, probably $15,000 to $40,000 for early deals, with either usage tiers by assets checked or workspace pricing plus policy modules (estimate). A services wrapper may be unavoidable at first, but a services-heavy model will drag founder time badly.
The minimum viable product should not try to be universal. The minimum viable product should solve one workflow for one kind of buyer: for example, a publisher ingesting mixed-origin images and text, or a regulated brand team routing creative through multiple generation tools. The minimum viable product should check supported signals, route exceptions, and keep evidence logs. The minimum viable product should not promise to prove authenticity.
The best time-to-truth test is the fast kill test: in 21 days, interview 25 named buyers across publishers, enterprise creative-operations teams, digital asset management owners, content management system owners, and trust-and-safety teams; show a one-page problem brief; ask for a paid 6- to 8-week pilot at $5,000 to $20,000 focused on one concrete multi-vendor workflow. Kill the idea if fewer than 3 buyers admit a live compliance gap and zero buyers commit budget, data access, or a dated pilot next step.
The risks are not subtle:
- incumbents absorb the feature;
- buyers stay inside one stack;
- detectors remain access-limited;
- labels create false positives and reputational blowback;
- the company becomes a slow consulting shop;
- regulations shift in interpretation;
- the founder spends a year becoming an unpaid educator.
Reasons not to pursue are equally plain. Do not pursue the business if the wedge is “we do watermarking.” Do not pursue the business if the founder needs quick feedback loops. Do not pursue the business if the founder has no warm route into publishers, trust-and-safety teams, or enterprise workflow owners.
The Full Top 3 Takeaways
1) The law created a real reason to buy, but the real reason to buy is narrower than “EU compliance”
The European Commission made the timing real by putting the transparency obligations into force from 2 August 2026. That date creates urgency. But urgency alone does not create a startup-shaped category. The actual buying trigger is not the existence of labels. The actual buying trigger is the painful moment when a company with mixed tools cannot explain, prove, or log what happened across a messy workflow.
2) OpenAI’s own caveat is the opportunity and the warning
OpenAI saying edited, translated, or short text may evade reliable detection is enormously important. The sentence proves the market pain is real. The sentence also proves that any startup depending on provider detection inherits technical fragility and trust risk. A founder should read that sentence as both invitation and hazard tape.
3) The only believable company here is a cross-vendor evidence layer
Adobe, Microsoft, Google, Meta, and OpenAI already own large pieces of provenance, watermarking, labeling, and verification. A startup does not beat those companies by pretending those pieces do not exist. A startup only has a shot if the startup becomes the neutral workflow layer that connects pieces, sets policy, and stores evidence where native tooling stops short.
The cliffhanger is simple and real: will large enterprises discover that native provenance tools are too fragmented and brittle for audits, or will Adobe, Microsoft, Google, Meta, and OpenAI close enough of the gap over the next 12 months that independent provenance middleware never becomes more than a feature?
Method: Super Fluidity Opportunity Dossier, public disclosure version. AI-assisted analysis and writing. Facts are source-verified where possible, with uncertainty kept visible where evidence is thin.
Why five ratings? Because the same idea is a different bet for different people. As a side gig — no money, no audience, no special skill, a few hours a week beside a job — the question is simple: is there a realistic path to $2,000–$5,000 a month within a year? As a business, starting with no money and no audience, most ideas are traps — even ideas that became famous companies rate only 2 stars from that starting line. With money, an audience, or both, the same idea can be a real opportunity. So every dossier rates all five, and you can find your own row.
Dossiers rated 0–2 stars for a founder with no money or audience are published free, right here. The rarer 3-, 4-, and 5-star finds go to paid subscribers first — that tier is opening soon, and free subscribers get the first invitation. Join the free list. Or have us investigate your own idea, privately: Request a Dossier ($199).
