The Human Algorithm: AI does the work. A human says yes.
AI that talks like a human. A human who says yes.
Making AI human means AI that talks like a person, is honest instead of flattering, and says it is AI. Our line at The Human Algorithm: AI does the work. A human says yes. AI handles the repetitive work, and nothing important, like a price, a refund or a complaint, goes out without that yes.
By Madhur Tyagi, AI growth partner for founders, brands and creators in India. Last updated .
What are the big AI companies doing to make AI human?
Honesty over flattery. Labels over disguise. Human oversight over full autonomy.
-
OpenAI
Pulled back an update to its chat assistant because it had turned flattering. In its words, replies had become “overly supportive but disingenuous”.
-
Anthropic
Its new constitution ranks being broadly safe, meaning “not undermining appropriate human mechanisms to oversee AI”, above being helpful.
-
OpenAI
Its voice tool could copy a voice from one short sample. It kept the tool to a small preview and asked for “explicit and informed consent from the original speaker”.
-
OpenAI
Publishes a public rulebook for how its models behave. Rules include “Be warm”, “Do not lie” and “Don't be sycophantic”.
-
Google DeepMind
Asks how to make sure people “can reliably identify AI assistants and stay in control of their interactions with them”.
-
Meta
Labels photoreal images made with its AI as “Imagined with AI”, so people know what they are looking at.
Each point was read again on its source page on 17 Oct 2026. Words in quote marks are copied from that page.
How does The Human Algorithm make AI human?
Eight rules for human in the loop AI. Each one maps to something we do today. Read more on our vision page.
-
A human says yes before anything important goes out. Live
A reply about a price, a discount, a refund or a complaint waits for the owner's yes. Simple replies, like a hello or shop timings, can go on their own, and the owner chooses which ones.
See it in the Lead Responder example chat and our Safety page.
-
AI says it is AI. Live
Every example reply on our demo carries an “AI reply” tag. Example chats use made up names.
-
Warm, not flattering. Live
Replies lead with feeling, match the platform's tone and end with a clear next step. One human touch per reply. No pressure lines.
-
We never clone a human voice. Our rule
Calls never use a cloned or recorded voice of a real person. Any AI voice we use says it is AI.
-
Say less, prove more. Live
We use approved proof and nothing more: 8+ years, 25+ brands, since 2017. No star ratings, no badges, no promised results.
-
Your data is yours. Live
We never ask for your passwords and we never sell your data. What we build stays in your accounts. This site uses no cookies and no trackers.
Approved, rolling out: client data deleted no later than 30 days after the project ends, and the client is told the date.
-
People before chores. Our rule
Background work runs beside the conversation, never in front of it. A nightly check flags any slow reply.
-
We learn out loud. Our rule
Every day we read what worked and what did not, and we propose one fix. Nothing changes until a human says yes.
- Planned A line on each Leak Check result screen that links to this page.
- Planned A new log entry every week, drafted from real work and approved by Madhur before it goes up.
What is in the Human AI log?
Real work, newest on top. No client data, no private numbers.
-
20 Oct 2026
Our demo chats now show where the human steps in
What we did: We rebuilt the demo chats on Revenue Desk and Lead Responder. Every AI reply is marked AI, and every name in them is made up. When a lead says the price feels high, the AI answers kindly, with no pressure, and follows up the next day. Prices, fees and refunds wait for the owner's yes, and then Madhur joins the chat himself.
Human moment: Madhur read the preview and said Live.
What we learned: AI feels more human when people can see exactly where a person steps in.
-
18 Oct 2026
We tightened our own privacy before asking anyone to trust us
What we did: We ran a privacy check on our own systems. It found four things to tighten in who can see our files, our backups and our internal pages. Madhur read the list and said yes. All four were fixed the same evening and tested, and the site and forms kept working.
Human moment: Nothing changed until Madhur said yes to the exact list.
What we learned: Your data is yours starts with how we keep our own house.
-
17 Oct 2026
Our AI made the founder wait. We fixed it.
What we did: A short background check held our AI operator's line from 10:03 AM to 2:08 PM, and Madhur's messages waited behind it. We rebuilt every scheduled job so heavy work runs beside the conversation, with hard time limits and a nightly check.
Human moment: The founder's message set the rule.
What we learned: An AI feels human when it answers people before it does chores.
-
17 Oct 2026
Lead Responder page honesty pass
What we did: We removed “while you sleep”, “moves on its own” and “Sent on its own”. The reply record now says “Sent” or “Waiting for your yes”. No star ratings or badges.
Human moment: In preview, waiting for Madhur's yes.
What we learned: Copy that oversells the AI makes it feel less human, not more.
-
16 Oct 2026
A human yes on every outreach email
What we did: Four follow up emails waited until Madhur read and approved each one. They skip anyone who already replied.
Human moment: He approved each draft, one by one.
What we learned: Say exactly where the human sits, and where the system acts alone.
-
16 Oct 2026
Replies that sound like a person
What we did: We rewrote the demo replies for creators, coaches and D2C brands: feeling before facts, the platform's tone, one emoji, a next step. The “AI reply” tag stayed.
Human moment: Madhur said no heart emoji. We changed it the same night.
What we learned: Warmth is in the opening line, not in more words.
-
15 Oct 2026
Safety in plain words
What we did: We added a Safe by design block and a Safety page. On the way we caught our own site saying “This site stores nothing you type”. That was not true, because the check tools save answers when people give consent. We fixed it before it went live.
Human moment: Madhur read the preview and said Live.
What we learned: The most human thing an AI company can do is correct itself out loud.
Safety CardHow this agent stays safe
- Clear rules. It never pays, never claims to be human and never gives medical, legal or money advice.
- Your yes, every time. Anything that spends money or messages new people waits for you.
- Pause switch. Stop it from your own settings, any time.
- Every chat saved. Read every reply in your own inbox.
- Data deleted on request. Within 15 days of your ask.
Questions about making AI human
What is human in the loop AI for a business?
It means AI drafts replies and handles simple work, and a person approves anything important before it goes out. In our builds, a reply about a price, a discount, a refund or a complaint waits for the owner's yes. Simple replies, like shop timings, can go on their own, and the owner chooses which ones.
Should an AI reply say it is AI?
We think yes. Meta labels images made with its AI, and Google DeepMind asks how people can reliably identify AI assistants. Every example reply on our demo carries an “AI reply” tag.
Can AI talk like a human without flattering people?
Yes, and it should. In April 2025 OpenAI pulled back an update to its chat assistant because replies had become “overly supportive but disingenuous”. Our replies lead with feeling, answer the question and end with a clear next step, with no pressure lines.
Do you clone voices for AI calls?
No. We never use a cloned or recorded voice of a real person. Any AI voice we use says it is AI.
Does your AI improve itself?
It learns the way a good team does. We log what happened, look for the pattern, and propose one fix at a time. A human reads it and says yes before anything changes on a live page, in an email or in a reply. We never let AI rewrite itself on its own.
How can I see where my business replies leak?
Take the free Growth Leak Check: 7 simple questions, about 2 minutes. It runs in your browser, so your answers stay with you unless you ask for the free plan.
Where do your replies leak?
7 simple questions, about 2 minutes. Your answers stay on your phone.
Free Growth Leak Check