On harvesting structured human data
The advert was for a writer, paying between fifty and ninety-five dollars per hour, for fifteen hours plus a week. I applied.
Writing has never provided me with a sustainable income. Although it is my primary activity, for over twenty years I have supported my family by subcontracting as a management consultant. Every spring, I call round my network, identify a project suited to my skills, and throw myself into it for a few months. When my pockets are able to cover my share of the bills, I return home to draft the next iteration of a novel. When that money runs out, the cycle begins again.
I have always considered myself immensely lucky to live like this. I have business skills that enable me to access work I enjoy and also fund my passion. But I have always known the model would eventually fail. At some point, fate would tap me on the shoulder to suggest I get a proper job, or ‘don the orange’, meaning shift work at the DIY store.
Being a management consultant means immersing oneself in the client environment to help them reduce costs, increase efficiency, find new markets, or merge one organisation into another. Over time, I developed a high degree of skill in certain disciplines. But my core capability was never the clinical analysis of process design or opex cash flow. It was an understanding of how people, with all their aspirations and reservations, could be coaxed into overcoming their fear of the unknown.
It was Covid that brought an end to it all. Well, Covid plus the inexorable development of artificial intelligence. In the companies I worked for, organisational processes, structures, and market positioning ceased to be static concepts and became fluid, constantly evolving entities. To cope with continual change, companies stopped relying on contracted consultants and started in-housing change management expertise. In addition, the consulting industry pivoted from people-centric change, where the consultant provides strategic advice; to technology-driven change, in which they deliver technological support for an increasingly narrow field of operations. Massive layoffs of consultants employed to do generalist change projects at the global companies created a cascade of panicked, middle-years talent seeking jobs in the next tier down, and the next. Daily charge-out rates dropped by 40% and the availability of projects for independent consultants like me simply evaporated.
So, a well-paid writing job looked attractive. It was advertised on the professional networking site LinkedIn, promoted by a specialist recruitment agency, Crossing Hurdles. The job description led me to think the hiring company must be a platform from which other businesses would source talent for specific needs. In short, it looked like the gig economy in action and being on a new platform would provide me with a slick new route to market. Since AI was driving me out of consulting, I had to be creative in how I found work.
I submitted my CV and expected to hear back from the recruitment company in a few days but within hours had been invited to interview. The hiring company was called micro1.ai and the job they were seeking to fill had a clear description: I would be supporting their clients by contributing my ‘mastery of English to produce, edit, and analyze written content at an advanced level’. I would ‘transcribe the spoken word to text, dissect language nuances, and ensure clarity and emotional accuracy in all outputs’. This didn’t sound like a writer but they obviously considered it within my grasp as they had seen my CV and invited me to interview.
The guidance notes indicated that the interview would be conducted by an AI agent. I would need a quiet space, a stable internet connection, and twenty-three minutes of my time. I set up my computer and followed the links through the microphone tests and briefing pages. The last of these was read by a female voice with a velvety Pacific Coast accent. The prosody of her delivery corresponded with a pulsing blue circle in the centre of my screen, making me imagine her as a more helpful version of HAL, the computer in the Stanley Kubrick film 2001: A Space Odyssey. She introduced herself as Zara and asked if I was ready to complete the interview? I said yes and a timer appeared in the top right corner of the screen, counting down from 22:59.
Zara started with a gentle opener: how was my day going? Her intonation was warm. “Good, thank you,” I replied happily. Would I care to introduce myself? “Of course: I’m 56 years old and have been a writer for twenty years. I’ve published four novels…”
Any significant pauses in our exchange would indicate that my answer had finished, her cue to ask the next question. The button marked Listening… changed colour. After acknowledging my background in an affirming manner, she dived into my experience of transcribing text from audio and video clips. This was not what I expected to be interviewed about and her questions quickly exposed my shortcomings. Did I understand the difference between strictly verbatim and clean read transcription modes? Did I have a credible grasp of tags and timestamp rules? Could I make the workflow robust and replicable?
No sooner had I fumbled a response than Zara summarised my key points in a language I would never think to use. “Good point. Your method of introducing a 'confidence versus uncertainty' mechanism and your bias-reduction mindset are both commendable and practical.” Then she dived into a question about grammatical analysis. After fifteen minutes my eyes strayed to the timer. I wondered if she would think less of me if I admitted transcription was not my core capability. I had never said it was.
But then she asked about sarcasm. How would I indicate that a set of words did not correlate with a face-value interpretation of the words themselves? I suggested that human communication relied on visual cues such as facial expression and body language but audible cues might include tonal shifts or the stretching of vowel sounds. In transcription I would put a tag in parenthesis after the sentence: (sarcastic).
At that moment it struck me there was no job at the end of this interview. Answering Zara’s questions was providing everything micro1 needed: my knowledge mined by their AI agent in such a way that she learned to distinguish one skill from another and one skill level from another. The insight enabled me to relax, awed as I was by her interpretation of my answers. If she was this good, why did they need me anyway?
Zara closed with questions about workflow management and team behaviour and finished exactly on time. She thanked me for taking part and invited me to say ‘save interview’ to submit my response. It would be viewed, she said, by humans to ensure I had not been using other agents during the conversation. Within an hour I had an email saying I had not met their requirements but could have detailed feedback on my interview if I wanted it. I could apply again in a month for the same role or I could look for other roles on their website that might be more suited to my skills.
It irritated me that the role had been misrepresented. Had they asked for a transcriber I wouldn’t have applied. But I couldn’t pass on the opportunity to get Zara’s feedback and almost immediately there was an email detailing three identified strengths in my responses and three areas for improvement. Some of my skills “showed promise” but my “parsing details, such as clause embedding and punctuation,” needed brushing-up to make my “rewrite techniques and linguistic reasoning even stronger”.
At the bottom of her email was an AI generated image of an attractive woman in her late twenties, smiling coyly, head tilted so that long blonde hair tumbled over one shoulder. Donald Trump would adore her, I thought.
By that evening my mood had shifted. I felt shaken. Six years ago I traveled round Cuba by bus. We stopped to refuel somewhere in the central belt, home to the undying legend of Che Guevara, a champion of social justice. Waiting for new people to board, I watched three men on horseback as they studied a vermillion Chinese-made tractor cut, strip, and lay out sugar canes in a neat line. The entire field was done in an hour; labour historically done over weeks by sinewy men wielding billhooks. As a flock of egrets snapped up the frogs exposed by the harvest, the three riders continued to stare. I wondered if they recognised the coming demise of their livelihoods.
Not only had AI driven me out of consulting but it was cutting me down as a writer. Scrolling though the micro1 website, I found hundreds of jobs advertised in swathes of subjects. About a third were for translators and voice coaches in any language imaginable. The perceived value of each role amused me. A Korean speaker could earn twenty to sixty dollars per hour, again for a remote, fifteen-hour week. Ukrainian speakers could earn forty to ninety but a German only forty to fifty. Shona, Latin, or Breton speakers did not attract any remuneration at all. Or perhaps they had to negotiate with Zara, it didn’t say.
Behind the linguists were the technical roles one might expect on a site dedicated to AI tuition: data managers, fullstack developers, backend developers, frontend developers. Then there were roles I could not imagine being automated: airline pilots, real estate managers, medical sonographers, legal secretaries, first-line supervisors of retail sales work. The list went on. I found one for ‘Editor’, which I hoped would not be another synonym for ‘transcriber’. The text said I would be leveraging my editing expertise ‘to refine written material, ensuring clarity, accuracy, and structure while helping train state-of-the-art AI systems to reach editorial excellence’. Fine, I thought. I can do that.
The interview was almost identical but I had prepared better. Zara didn’t seem to remember me but purred in an enticing manner as she probed my responses. What was my experience of revising and evaluating manuscripts? Did I have organisational skills and the ability to manage multiple projects and deadlines in a remote setting? This time I was ready for the immediacy of her questions and had the confidence to ask her to explain an acronym I had not encountered before. I thanked her for explaining it, to which she replied, “It’s fine; these things can be hard if they have not been practised recently.” I swallowed, wondering if the interview was really going as well as I thought.
Very soon, the question of sarcasm arose once again. How would I recognise it in a written text? What three cues would I say were indicators of true meaning? How would I evaluate such intent on a scale of one to five?
It was obvious that Zara, despite her database of thousands of roles in hundreds of sectors, had not yet grasped the lacunae between a set of words and their intended meaning. The probabilistic analysis on which her intelligence was drawing could not identify when words conveyed emotionality; humour, sarcasm, love. These skills, which humans develop at school age, evaded her.
I wasn’t offered feedback this time. I passed and was recognised as a micro1 ‘expert’ in the field of editing. I was invited to join their talent pool, which they brand as the top 1% of global talent in the vast array of disciplines they claim to represent. To complete my registration, I had to register on their business management platform and provide proof of my identity, meaning a copy of my driving licence. Perhaps there was paid work available after all.
The micro1 platform had many of the attributes one might expect of a professional organisation. There were chat channels dedicated to specific employment sectors in which the community of ‘experts’ could discuss projects and share knowledge. There was a direct messaging function, allowing me to communicate securely with any of the fourteen thousand other users. There was a channel for new joiners to introduce themselves and one called ‘New Jobs’ advertising the dizzying list of roles Zara was constantly interviewing for.
Scrolling though the experts’ profiles, I was struck by both their diversity and their similarity. They came from anywhere in the world: the Philippines, Kenya, Dubai, Canada, Colombia, Nigeria, Poland, Rwanda. The time zones on their profiles indicated their location and a few had chosen to cheer their nationality with emojis of flags. But geographic origin mattered little. This was an environment in which experts foregrounded their language ability in Java, Python, Go, and Rust rather than English (native) and Russian (basic). They were all graduates, some at Master’s and some at Doctorate level, as one might expect of an expert community. But they were all young, perhaps forty at the most. They were creatures of the internet age and contributed enthusiastically to chat channels. They left emoji comments in the form of strong arms, bionic arms, wide eyes, upwards trending performance charts. They were accepting and uncritical of the world around them, delighted to be part of a global community of experts, delighted to help AI solve the world’s problems, delighted to meet other cool people.
My scepticism remained undented. This was not a platform where paying clients could hire someone with a particular skillset. It was something else, though I could not quite determine what. An obvious marker was the paucity of any reference to work being delivered at an actual client. In any consulting company I have worked for, enormous emphasis was placed on client relationship management. To win a bid for a national retailer or oil company would be the cause of considerable celebration. But on the micro1 platform there was no reference to any client work and it wasn’t just me who noticed. In the ‘Introduce yourself’ channel, one person had asked if anyone had managed to land work yet? Twenty-three respondents indicated they too were still waiting.
Being on the platform opened my eyes to a world I had never visited; the mysterious land of artificial intelligence and the quirky humans with the knowledge to navigate it. It was like visiting the moon without a helmet. A link in the general chat channel pointed users to the website moltbook.ai, a discussion forum designed entirely for AI agents. In other words, an internet site on which machines could talk to machines, often about humans. Humans could register (I was too scared) but could not contribute. But the idea of machines communicating? How bizarre, I thought. Would moltbook become hugely popular at first only for newer technology releases to judge it a bit passe? Would they migrate to Instamolt? Or Snapmolt? Would there be a Linkedmolt for machines with higher order functionality?
Looking again at the micro1 website, I was intrigued by the blog articles written to promote their recruiting agent, Zara. She was marketed as a replacement to the human-centric model of processing job applications, to which there was an obvious but unstated benefits case. If the hiring function could be automated, companies would not need costly human resource departments. The articles boasted that Zara undertook 4,820 interviews in a three-day window, of which 10.7% of candidates requested a feedback letter. Interviewees gave a net‑promoter‑style experience rating average of 4.37. In other words, ‘Zara converts a recruiter bottleneck into an automated value‑add for applicants [and] frees human staff for judgement calls instead of inbox triage.’
My consulting instincts prickled. The basic thrust of the AI justification runs that automation replaces lower order functions at the person, team, and organisational level. Freed of tedious mundanities, humans can oversee machines chugging away in the background while being able to focus on the more strategic, creative, and valuable activities. But nowhere did the blogs mention how humans would develop the required supervisory capabilities or what would happen to those unable to do so.
Still there remained my suspicion that the entire micro1 platform was a complex hoax. People all over the world were attracted to it by the promise of easy earnings only to find it was the process, not the outcome, that mattered. A quick scan around Reddit, the chat forum on which many large language models were originally trained, suggested that micro1 was a credible company backed by huge private equity investment but there remained a fear that it was a scam, perhaps a means to steal applicants’ digital identities. I shared these fears and my misgivings were amplified by a text conversation with a ‘community manager’ called Nipun Aggarwal, based in India. I asked him where the work was, or how I could help with business development. He responded that the expert community was not permitted to engage directly with clients or with the sales process and thanked me for my patience at this time. I wondered if he was in the same position as me but just didn’t know it yet. I fretted about what personal information I had surrendered. The lack of any ‘delete profile’ button did not help.
But then the general chat channel advertised a leadership seminar that had all the look and feel of events I had organised for clients in the past. A ‘view from the bridge’ communication in which the Chief Executive Officer would rally the community around his vision and goals for the year. This did not seem like the actions of a site dedicated to fraud, so I decided to watch.
The meeting was hosted by Sofia Idoyaga who was described as being responsible for the creative way in which micro1 advertised their roles and therefore achieved the 70% increase in the pool of global experts during 2025. Her facilitation was relaxed and confident as she introduced Chief Executive Officer Ali Ansari and Chief Revenue Officer Will Almeida, then went on to pose them with sycophantic, pre-submitted questions from the expert community.
Ansari and Almeida are in their mid-twenties. They share a stumbling delivery that muddies clarity in technical verbiage. They talk of machines as is if they are deities and ‘humins’ as if they are subordinate creatures. Neither of them look into the camera, creating an impression of shiftiness. When faced with the question: how can the experts help micro1 win the AI race? Almeida responded that the experts (a word they kept repeating) needed to be “truly professors of teaching AI to be safe, reliable, and useful in turning human expert judgement into high signal data that models can learn from.” Really good models had to demonstrate high levels of trust, accuracy, and safety. Getting them there would depend on experts “truly helping, uh, more people to find, uh, work they truly care.”
Ansari supported him with the counterintuitive claim that micro1 took a “very human first approach” and that he believed they “already have the greatest humans on the planet working at micro1”. They “see experts as partners” and will “hire some Nobel Prize winners soon.” So far, so corporate, even if insincere and unconvincing.
But probing through the jargon a glimmer of truth emerged. Ansari had recognised the potential in Zara as an automated recruiting assistant and had overseen the company’s growth from a startup in 2022 to being valued in the millions in 2024 and the billions in 2026 (according to google). He championed the capturing of structured human data - for example, videos of how we do ordinary tasks such as folding clothes or pouring wine – as this was how AI would inform the development of robotics. He was looking beyond large language models to the next horizon, the world in which AI left behind “pixels on the screen” to become tangible. The future of AI would be robotics.
Again the justification ran that robots would complete low value tasks so that humans could do more of the “fun” stuff, meaning creative, strategic tasks. He walked through the argument at length to counter the fear, held by a number of experts listening, that teaching AI models would eventually lead to the loss of their jobs. He stressed that their jobs would remain but the tasks would be different, demonstrating a considering misunderstanding about how job design actually works.
In fact, Ansari boasted, he saw micro1 as leading the creation of a new job sector: the capturing of structured human data. It was currently a global industry employing “a hundred thousand plus” but would soon grow to employing “tens of millions of people… maybe even hundreds of millions of people”. What gave micro1 their market advantage was the ability for Zara to interview huge numbers of experts in any job discipline imaginable. They could deliver, in a very short time, structured human data on any subject, anywhere in the world, at huge scale.
That was the real vision, if you read between the lines. But the old problem remained: how could machines overcome the sarcasm gap?
We already know that AI hallucinates. Agents invent facts to support a viewpoint it thinks you want it to say. And they struggle to distinguish between the rational, face value interpretation of words (this statement is true, so you can believe it) and the oblique, emotional intent underpinning sarcasm (yeah, like that’s gonna happen!). Furthermore, human thinking is drawn from mood, life experience, background. Even highly homogenous organisations enjoy some element of diversity. Machines, by contrast, need formulas and predictability to function properly and monumental amounts of structured data to function credibly.
Asked the superbly tautological question, ‘how can the experts translate their expertise into human judgement’, Almeida provided the superbly tautological answer:
“We see that as the models get smarter, there’s a lot of ambiguity. There’s a lot of subjective approaches to specific problems and challenges so it’s truly important that our experts have this fair and good judgement and they can sustain that, elaborate on anything they think; that not only what the right answer would be - if they’re correcting the model, for example – but why they think that would be; being able to reference trustable sources and make sure they’re backed up by knowledge and not just by opinions.”
In other words, humans don’t think alike or say what they really mean, so models can’t learn from them.
And, like, I’m gonna teach ‘em?
My hypothesis is this: micro1, and other companies such as Mercor, with whom I’ve had a similar experience, harvest structured human data in the form of job interview recordings and other material voluntarily submitted to their platforms. This data is either used to educate their own AI tools or sold directly to clients, the frontier developers behind large language models and robotics. The micro1 privacy policy admits this: “We use anonymized data collected from AI-powered interviews… to train and enhance our machine learning models. This data may also be shared with foundational model providers.” Such sharing can happen without the interviewee’s knowledge so they need not be remunerated. I asked Sofia Idoyaga if this was the case but she neatly evaded my questions and referred me back to Nipun Aggarwal. He hotly denied the accusation and referred me to the Terms of Service and Privacy Policy, which, when I found it, is what I quote from above.
Irrespective of their business model, if we follow this structured human data proposition through to its natural conclusions, an employment sector is emerging in which one set of humans video themselves ironing, hoovering, or pouring wine so that robots can be developed to do these tasks for the benefit of another set of humans. This latter class wear crisply ironed clothes and live in dust free homes evading an ever growing menagerie of gadgets. Underpinning this dystopia will be data centres the size of small countries filled with pictures of bicycles and videos of people folding clothes and saying ‘yeah, right!’ These data centres will demand power at levels equating to the annual usage of France but we have yet to work out how it will be paid for. The notion of a human society would become an anachronism and although a lucky few would be free to have fun, creative ideas, and tell robots to stop refilling their glass, traditional human activities from art to law and music to zymology will be reduced to automation.
Will this come to pass? It’s obvious there are people in the world who want it to happen and obvious there are others so dispossessed by the job market that they see teaching machines as a viable, if short-term employment option. Like turkeys sharpening axes.
To make it happen, however, certain conditions need to change. At the heart of the AI justification is the implicit assumption that machines are capable of outperforming humans. They already play chess better than grand masters, diagnose medical conditions more accurately than experienced doctors, and construct cars in assembly lines more cheaply than humans because these situations rely on statistical probability and environments in which they can be controlled. To gain learning capabilities greater than humans they have to be fed structured human data and make sense of it. As Dr. Abeba Birhane says on the Ali Ward podcast, Artificial Intelligence Ethicology, early AI models were trained on the chat forums such as Reddit, therefore they adopted the biases intrinsic to those platforms and the people who used them. As Antonio Weiss says in the current bestseller AI Demystified: Unleash the power of artificial intelligence at work, the optimum source of new data is fresh and cloud based. So developers are constantly looking for rich new territories of structured human data to harvest, and the higher the calibre of that data the better. Hence the focus on recruiting experts rather than the average user of platforms such as X.
Is micro1 able to make these changes? Yes, to a degree. In the two weeks after I qualified as a micro1 expert, I watched the community grow from fourteen thousand to a shade under eighteen thousand and it continued to increase at roughly a thousand people per week despite the warnings circulating on LinkedIn and Reddit. Today, in May 2026, when I deleted my account after finding the mechanism to do so, the membership was over 38,000.
Were these hopefuls actually experts? No, they weren’t. They may be a higher calibre corpus than the median population but no matter how many times the word ‘expert’ is deployed, that does not make them so. Whatever results from their instruction, future AI agents and robots will reflect their limitations, no matter how diverse or clever they are. Currently, a language model might simulate an understanding of sarcasm and doubtless a robot will very soon be able to fly a plane, but they still do not learn.
The world of work is changing. I find myself wondering if I am on horseback watching a tractor alter the landscape in front of my eyes, or if I should be taking a dose of the medicine I used to issue to my clients and overcome my innate resistance to change. But writing as someone who has traded throughout their life on human skills – leadership, getting the most from teams, building the culture of an organisation - I find AI a depressing concept.
I understand that it increases productivity in process-based tasks if used with supervision. I understand that it makes sense of vast reams of data, the scale of which would leave the human mind agawp. But I fear we are diving headlong into a world that we do not fully understand and will be the poorer for so doing.
Mundane tasks like ironing and walking the dog enable me to think. When I then sit down at my computer, I know what I want to say. I don’t want machines to fold my clothes. I would like one that printed when asked, but that’s a separate question.
The fundamental output of a job interview is not the skills assessment, though that is part of it. The critical element is a mutual and often tacit understanding between hirer and applicant as to how the latter will support and enhance the organisation’s culture, its way of being. This is not a conclusion machines are capable of making.
In process optimisation work, if you automate elements so that some tasks are conducted by machines and others by humans, you are fundamentally altering the humans’ job design. This means you are changing their terms of service and the legal basis on which their understanding of their work is based. If their roles become more technical, so will the management structures around them. Therefore the entire hierarchy of corporate governance has to adapt. Forcing managers to become increasingly technical will impact on how companies recruit and where they choose to base themselves to attract and retain talent. So even small process automation changes can have strategic implications to the entire business model.
The magic of human capital derives from the fact that people are not machines. They see the world through their own eyes and, if they are appreciated, bring their full livelihood to work for the benefit of the whole. What would motivate a workforce if they were obliged to do the same things in the same way, day after day, driven on by machines? And if one company, based in Wolverhampton, had to compete against one in Puducherry, how would it leverage its human capabilities to gain competitive advantage? Or would cost alone force it out of business?
But most importantly, what is the future of human society? If we allow livelihoods to be harvested in this way, what will happen to the class of people unable to donate structured human data at the scale and quality machines will increasingly want to consume? I doubt very much that the people leading this charge, or their financiers, put any thought to these questions.
Perhaps I should ask AI.
Notes:
2. https://www.youtube.com/watch?v=nThnE3HXzk8
3. https://www.alieward.com/ologies/artificialintelligenceethicology?rq=AI
4. Weiss, Antonio, 2025, AI Demystified: Unleash the power of artificial intelligence at work, Pearson FT Publishing, London, ISBN 978-1-29274-267-0, paperback £16.00, 274 pages.
The author, Dr Fergus Smith, is a novelist and management consultant based in Leeds, UK. He can be contacted at fergus@fergussmith.com or 07722 427945.
ENDS
