When the record of who caused something passes from one company's systems to another, it does not survive the trip, so everything measured after that is an estimate. Advertising has lived with that for twenty years, so it is a good place to watch it happen.
Machines have been transacting with each other, at scale, with no human in the loop, for twenty years, and almost every ad you see online was bought this way, in an auction that ran while the page was loading. The industry calls it programmatic advertising, and it clears billions of those auctions a day. Each auction settles in roughly a tenth of a second, money moving between parties that have never met and never will, at a speed where nobody could be in the room.
It runs on an identity layer that has never quite worked.
A cookie is a small file a site leaves in your browser, and for years it was how anyone knew a click and a purchase came from the same person. Browsers started removing them, then partly walked that back, and the ad industry built workarounds either way. What replaced it was reading the combination of details a browser already leaks on its own, screen size, fonts, the timezone it's set to, and matching a person to that combination instead, which the UK's privacy regulator called irresponsible when Google opened it to advertisers.
What is left is the attribution window, the stretch of time a publisher gets credit for a sale after someone clicks. That one is not a measurement. It is agreed between the publisher and the platform, and the number they agree on decides who gets paid; if you have ever watched two ad platforms report different totals for the same campaign, this is why. Neither side can prove the other is wrong, and that argument has been running for twenty years.
Nobody fixed that. What got built instead is a second industry that brands pay to check the first one's work after the fact, asking whether a real browser loaded an ad or a script pretending to be one, whether it was actually on screen rather than sitting loaded in a tab nobody looked at, and what it landed next to.
All of that runs after the auction has already settled and the money has already moved, so none of it stops a bad transaction. It turns into a refund, or a worse score for that seller next quarter.
Grading works when the thing being faked is traffic, because you can look at a request and estimate how likely a person is behind it, and it stops working when the fake thing is the person.
Where the record gets lost
Affiliate marketing is the simplest version of it, so it is worth walking through slowly. If you follow it here, you have the whole idea. A creator posts a link. Someone clicks it, buys something, and the creator gets a cut. Three things track that sale, the link itself, a cookie dropped in the buyer's browser, and a spreadsheet or dashboard somebody maintains.
All three live inside one company's systems. The moment the sale moves to a second company, a payment processor, a different retailer, an ad platform, the record of who caused it does not travel with it. The second company starts its own record from scratch, and it has no way to check the first one.
Now two companies hold two records of the same sale. Each one prefers its own. There is no shared place to check which is right, which is the publisher-versus-platform argument from the top of this piece, showing up again in a different industry.
Nothing about it is specific to advertising. It happens wherever two systems hand something off without a database in common, and if you are building anything that passes a record to a partner, you have this. The thing being handed off can be a referral, a payment, or a claim about who somebody is. The record does not survive the handoff, so whatever gets built on top of it is working from an estimate.
When the fake thing is the person
In 2023 AI video was a grainy Will Smith eating spaghetti clip everyone passed around, and in June 2026 a fully AI-generated feature premiered at Tribeca. 75 minutes, about two thousand dollars, made in London by a director who couldn't get into Iran, with every person on screen generated.
The same tools clone a voice.
The FBI broke that out separately in its 2025 Internet Crime Report, giving AI its own section for the first time in the report's nearly 25-year history. 22,364 complaints referencing AI, 893 million dollars in losses. Voice cloning of relatives and executives, deepfaked endorsements, synthetic identities propping up fake investment communities.
The wider shape is impersonation, with or without AI. The FTC's 2025 numbers put imposter scams at 3.5 billion dollars across nearly one in three fraud reports, more than any other fraud category that year, and the fifth straight year it led. Nobody breaks out how much of it is synthetic, so the 893 million counts only the part somebody flagged.
When nobody can prove where an answer came from
AI Overviews are the answers Google writes at the top of a results page, using content from the sites below it. Penske Media is suing Google over them, and the complaint lays out the old arrangement plainly. Publishers let Google crawl their content, and Google sent readers back. It then quotes a Google scientist observing that direct answers "reduce search referral traffic," and the complaint ties that to lost publisher revenue. The research it cites puts click-through on the top organic result down 34.5 percent.
Then in late May 2026 a court in Munich ruled Google couldn't claim host-provider protection for inaccurate AI Overviews. That protection, written into the EU's Digital Services Act, is the rule that a platform is not liable for what other people post through it. Google said in June 2026 it would appeal, so none of it is settled law yet.
Publishers are now paying to have their material rewritten so that machines answering questions will quote it correctly, whether or not a reader ever clicks through. The work has a name, answer engine optimization, and every one of those projects runs into attribution, which stays open because no party in the chain can prove what a given answer was assembled from.
Licensed voices, and where the license stops
The consenting version of this exists and has names on it. In November 2025 ElevenLabs launched an Iconic Voice Marketplace, and Michael Caine, then 92, licensed his voice into it. Matthew McConaughey signed separately, invested in the company, and now uses his own clone to publish a Spanish-language version of his newsletter. The catalog also carries people who are dead, licensed through their estates, including Judy Garland, John Wayne, Maya Angelou, Alan Turing and Babe Ruth.
In May 2026, ElevenLabs licensed Stan Lee's voice and likeness from the venture that manages them, seven and a half years after he died. Val Kilmer's estate cleared his image and voice for a film called As Deep as the Grave, where a generated version of him leads the picture after his 2025 death.
Tennessee passed a law in 2024 treating a voice as part of a person's identity rather than a feature of their celebrity, and California and New York have widened what an estate can control after death.
I have my own voice clone on ElevenLabs too.
A company submits a request for a voice, and ElevenLabs matches that request to whoever holds the rights. The two parties then negotiate off-platform, and once they have agreed, ElevenLabs generates the audio.
Everything difficult sits in that off-platform step. The software introduces two parties and hands them a contract to sign somewhere else. Nothing in the public process describes a record tying a particular generated file to a particular license, or a payment when that file plays, so the marketplace solves finding the rights holder.
Transcription has the same boundary, and there a company states plainly that it chose it. AssemblyAI sells speech-to-text infrastructure and will separate who spoke in a recording, then put names on the segments by reading the conversation for them. It does not keep a voiceprint. Their own writing calls that a deliberate design choice rather than a gap, with the embeddings computed in memory and discarded per file so there is no biometric database to manage. Picking a specific known person out of a group by their voice means storing enrolled voiceprints, and they describe the privacy and compliance weight that carries as a reason not to.
A shared record, still in draft
A draft Ethereum standard, ERC-8004, proposes three registries, shared lists that no single company owns, which is the part adtech never had. They hold an agent's identity, a record of how it has behaved, and outside checks of its work, and the draft is still changing, with authors at MetaMask, the Ethereum Foundation, Google, and Coinbase.
Why I care, and what I've built toward it
I grew up around people getting taken advantage of. My dad got scammed out of money from different ends, and it affected our family.
My dad was a man of few words. He showed you by taking action, and most of that action was work no one sees. He showed me what integrity was, and he gave foreign students the opportunity to succeed in this country by furthering their own education.
He also had a habit of taking on more than he could handle, and it took a toll on him. He got sick in 2001 and again in 2008, and each time what I was given was agency, inspiration, and the motivation to build something toward a better outcome. It also grounded me, because I took on more responsibility early by looking at the full macro environment with the tools I had on the micro level, and had to become more resourceful and more creative to handle it.
What he passed along was the opportunity of continued education, and those were the seeds that were planted to always stay curious, when no one was around. Every summer was some kind of academic camp, including iD Tech Academies at Montclair State University for robotics, applied mathematics, and visual arts, and every day had a workbook. He would drop my brother and me off at Barnes & Noble, which was like a daycare, and that is where I was reading all the time.
Where my career started
Those habits carried into a federal work-study job at NJIT in 2008, putting professors' course material online, where my career started. What that job gave me was being able to take my agency into a structured environment, with a culture and people around me that allowed me to learn more, build solutions, and solve problems, which was the real value. I also spent more time learning what I wanted to learn on Wikipedia. Wikipedia attributes dates, timelines, and history, and history has a tendency to repeat and shape trends, which is what you see when you work on the fintech side.
That same year my dad lost his business. It was another time when being resourceful and creative, and owning the outcome, mattered, and it is how I later got the opportunity at Zeta Global.
From there my career ran through adtech and fintech, on both the creative side and the engineering side. For me it was about the principle. I spent my whole career in these industries, and the question I kept asking was what all of this meant for the people on the receiving end of AI. The people I met along the way gave me the opportunity to share pieces of my story, and that let me continue this work.
What I'm building now
Now I'm using every tool my dad gave me, together with AI, to build toward a safer world for people. AI feels like the modern version of walking into that Barnes & Noble, except now the library talks back. It explains, it helps you build, and it can cite sources, and I have also seen how systems like it get designed for engagement instead of understanding.
I've been working in Runway, cloning myself. What I'm building is templates. It's how I've built everything across my life and the places I've worked, where the workflows and programmatic work were templates too: everything is a template. The skincare ad on my homepage is one of them, a product ad with a generated influencer, made once and localized into English, Japanese, and Korean, and each card labels her as an AI influencer.
In March 2024 I left Interactive Brokers and Chicago, a city I had lived in for two years that gave me life experience and more learning, because being close to home and family mattered more. I took the time to focus on things that mattered more in my life. That April I started Trending Society as the framework I could build on.
Starting in April 2026, I built its article pipeline, which checks every draft against 10 structural AEO rules and 7 judge criteria and rewrites any draft that fails.
The pipeline lives in a private Trending Society repository, so here are six of its article tasks and the two check files, by name only:
packages/trigger/src/
├── tasks/article/
│ ├── single-article-pipeline.ts orchestrates one article, end to end
│ ├── classify.ts classifies and filters the source
│ ├── verify-claims.ts checks the draft's factual claims
│ ├── article-rewrite.ts rewrites a draft that fails a check
│ ├── generate-narration-script.ts
│ └── sync-article-to-notion.ts hands the draft to the Notion CMS
└── lib/content/
├── aeo-validation.ts the 10 structural rules
├── llm-judge.ts the 7 judge criteria
└── 11 test files alongside the checksHow I work
I have always built with a mission, and the places where I did my best work had strong missions, culture, and team structure, where the work was a reflection of them. Building for the customer started in high school at Staples. I learned where every item in the store was, so I could tell any customer exactly where to find it. I always felt that walking them to the item, and understanding more about why they were there and more about the person, before everything became digital, was nice. On one slow, rainy day I reorganized the leftover back-to-school binders without being asked. The store earned strong customer-service rankings, I made employee of the month, and the lesson I took was to do the quiet work and dedicate the time, because the work speaks for itself.
Years later on Champions of Cash, I built the desktop unit and Jonathan McGough built mobile, and reviews ran on working builds, so the ticket stayed tappable and the disclaimer stayed clear before anything reached the client.
Staples, Champions of Cash, and the Trending Society pipeline are the same method at three sizes. Learn the whole system before changing one part of it, turn the repeatable work into templates, and put written checks in front of anything that ships. Because it is written down, the method goes with me to whatever team I join, and the team can run it too.
What my life has taught me is to be persistent and put in the hard work no one sees, the same work my dad did. What feels random is usually a compilation of intentional decisions, and those decisions get tested along the way. Waking up to tackle problems like these is what excites me, which is why I am particular about where I plant my feet and root in, somewhere I can bring all the work and learnings to a team that can do more with them. AI is only as good as the people steering it, and the feedback loop gets better with other people's perspectives on how they use it, so sharing those learnings as a team goes a lot further than working solo.
Why First Day
Without a label, a viewer watching a generated ad has no way to know it is generated. The people who will spend the most years inside that arrangement are teenagers now, which is why First Day is an architecture decision for me. I treat it the way I keep architecture decision records, with the reasons and principles written down: a better story for the younger generation, and a better place for them to consume content. What the right disclosure looks like, and who should define it, is still an open question across the industry, and building the creative and the engineering under it keeps me close to that question.
Starting with First Day meant starting with the hardest problem. Kids are an audience where verification and safeguards are a core requirement, so I won't deliver something to them without the right architecture, one whose output is deterministic. Building for that audience puts the safeguards into the architecture from the start, instead of adding them to UGC ads later, and it keeps me asking who the audience is and who is inheriting the solutions and inventions today.
High-stakes products like that are the challenge that excites me, and the skills and the harness I have built, the written rules and checks my AI agents run before anything ships, are the foundation for them. It is a long road that takes a team, and everyone raising a family or building these tools today will look back on whether we did our best job delivering it.
In June 2025 I started researching what would change for students if federal education funding went away, to stay close to the problem, and that research became the thesis behind the AI tutors I build with the Runway platform as part of my AI research and creative lab.
In August 2025, through Trending Society, I started a pro bono internship, posted on LinkedIn, so students could learn by building alongside me, the same seeds my dad planted, applied. It grew into more than one person could run alongside everything else, and it taught me that people need to learn agency on their own, the way the work-study taught it to me, so I wound it down and put that time back into building.
In June 2026 Runway accepted me into its Builders program, one of 65 picked, and First Day is built there.
I can't stop this from happening, but I can put the safeguards in. Building with AI, I hold myself to operating with integrity for people like my dad, or anyone else, whatever their background or age, so that whatever is inevitable, people get a fair share and a voice in the outcome.
See the work
- First Day, where the architecture decision above gets built
- Champions of Cash and Garmin vívoactive, the adtech builds where the method started, with the collaboration and teamwork behind it