The story and the people behind Honest Lens

Summary

Every day we receive dozens of messages, headlines and “breaking news” posts with no clear origin. More and more people turn to AI chatbots to check them. But chatbots tend to answer with complete confidence even when they have no evidence, and sometimes they cite sources or links that do not exist at all. A tool meant to stop misinformation can quietly become a new source of it.

Honest Lens is our answer to that problem. It splits any news text into separate, checkable claims, weighs each claim against evidence, actually opens and verifies every link it cites, and when the evidence is not there, it says “Insufficient Evidence” instead of guessing. It has even measured its own accuracy in public and published every result, including the cases it got wrong.

The problem: why this tool was needed

1. A flood of information and forwarded messages

Much of today’s news reaches people not from news agencies but from messenger groups and channels: a message quoted “from a reliable source”, a number stripped of its context, a headline that has drifted far from the story it came from. Most people have neither the time nor the tools to trace each one back to its origin.

2. Chatbots were not built for fact-checking

General-purpose chatbots are remarkable tools, but they have built-in weaknesses as fact-checkers. They give one overall answer for the whole text, rarely show which sources they actually checked, may produce broken or invented links, and, most importantly, seldom say “I don’t know”. None of them tells you how accurate it is at fact-checking.

3. A gap for Russian and Persian speakers

Most serious fact-checking tools are built for English-speaking audiences. Russian and Persian speakers either work in a language that is not their own or have no tool at all. Persian interfaces in particular often implement right-to-left layout only halfway: text breaks, numbers jump around, and the experience becomes frustrating.

4. A crisis of trust

People do not trust a tool that simply says “true” or “false”, and they are right not to. Trust is built when you can see for yourself what a verdict rests on, and when a tool does not hide its limits. That is the kind of tool we set out to build.

The idea: the philosophy of “I don’t know”

The project started from one sentence that later became its slogan and the compass for every decision:

“A fact-checker that doesn’t pass absolute judgment, but shows evidence; and when it doesn’t know, it honestly says ‘I don’t know’.”

In practice, this sentence shaped several key product decisions. First, instead of a 0–100 “truth score” that looks scientific but explains nothing, each claim receives one of four clear verdicts, together with its reasoning and sources. Second, “Insufficient Evidence” is a complete and respectable answer, not a system failure. Third, the tool has to be as honest about itself as it is about the news, which means measuring its own accuracy and publishing it.

The solution: three pillars

Honest Lens is built on three pillars, each answering a different part of the problem.

Pillar A: the claim-to-evidence engine

Users paste anything: a headline, a forwarded message, a whole article or just its link. The system reads the text, extracts the checkable claims and examines each one separately. For every claim, the output includes a verdict, an explanation, a confidence level and sources. Splitting the text into claims matters, because a news story is usually a mix of true, half-true and false parts, and a single answer for the whole text would itself be misleading.

Pillar B: the Distortion Tracker

Sometimes a story is not false, just distorted: a number has grown, a qualifier has disappeared, or a sensational tone has been added. The Distortion Tracker places a media report next to its original source and shows exactly what changed: added spin, altered numbers and missing context. It is useful for journalists, researchers and any careful reader who wants to see the gap between what was said and what was published.

Pillar C: the public Self-Audit

This is where Honest Lens truly differs from similar tools. We selected a set of real claims from different fields, established a reference verdict for each from public sources, and then ran all of them through the real engine. The results are published without any editing, including every disagreement. The details of this test are below.

The user experience: from a suspicious message to a shareable report

We wanted checking a story to feel as simple as forwarding a message. The user journey has three steps:

1. Paste it. A headline, a forwarded message, a whole article or its link. No sign-up required.
2. Watch the check live. A deep analysis takes a little time. Instead of an empty loading screen, users see what the system is doing as it happens: the article was read, context was looked up in reference sources, claims were weighed against evidence, and each link was verified. The wait became part of the experience and a source of trust.
3. Get an answer for every claim. Each claim has its own verdict, explanation and sources. Every report also has a permanent link that can be sent back to the very group the story came from.

That permanent link is a small detail with a big effect. Misinformation spreads in group chats, so the correction has to be able to spread there too. Instead of an argument, the user shares a link that shows the evidence.

Real link verification

One of the features we are proudest of is that Honest Lens never trusts a link blindly. Every link cited in a report is actually opened and checked. Broken or invented links are removed, and the user is told that they were removed. As a result, users never meet a source that does not exist, a problem that is common in chatbot answers.

Why not just ask a chatbot?

Honest Lens uses AI models too. The difference is everything built around them. The difference at a glance:

The Self-Audit: being honest about ourselves

The easy option was to write “99% accurate” on the website. Many products do exactly that. But a tool whose slogan is honesty cannot publish invented numbers about itself. So we designed a real test.

• 36 real claims were selected from five fields: geopolitics, economy and trade, science and health, historical records, and media sensationalism. The claims came in all three languages: English, Russian and Persian.
• For each claim, a reference verdict was established from public, citable sources, with the source link shown next to it.
• All claims were run through the real engine. No result was written or corrected by hand.
• The outcome: 75% agreement with the reference verdicts, 19% honest “I don’t know” and 6% disagreement.
• All 36 cases, including every disagreement, are published on the Self-Audit page with the engine’s full report and the reference source, so anyone can check them.

The most telling number is the 19% of “I don’t know” answers. In those cases, instead of issuing a wrong verdict, the engine stated that it did not have enough evidence. That is exactly the behavior we designed for. Real errors, verdicts that contradict the reference, were only 6%, and even those are not hidden. We believe that publishing its own mistakes is the most valuable proof of credibility a fact-checking tool can offer.

Design and visual experience

An identity that conveys trust

For a product about truth, the design had to be calm, clear and serious rather than flashy. The Honest Lens identity is built on deep cobalt blue, white and a single “signal” dot of color that also appears in the logo. We explored several design directions and chose the one that felt professional and journalistic while working equally well in all three languages. The banner, poster and social media preview images were designed in the same visual language.

Three languages, no compromises

Supporting Persian was not a simple translation for us. The entire interface was built from the start with right-to-left logic, so when the language changes, the layout, icon direction, spacing and tables all mirror correctly. Numbers are shown in each language’s own format, and fonts were chosen separately for the readability of Persian, Cyrillic and Latin scripts. Each language also has its own address, such as honestlens.ir/ru and honestlens.ir/fa, so users and search engines see each version on its own.

A home page that explains the “why”

In early versions we noticed that the slogan alone was not clear enough to first-time visitors. We redesigned the home page to answer the questions in a visitor’s mind in order: what this is, how it works, what verdicts it gives, how it differs from a chatbot, and which sources it relies on. A moving ticker shows sources that the engine has actually cited in its reports. We deliberately did not use media outlet logos, because we did not want to imply an endorsement that does not exist. Finally, an FAQ section answers common concerns: Is it free? Can it be wrong? What happens to the text I paste?

For everyone, on any device

Most people receive news on their phones, so the design was mobile-first from the beginning. Color contrast, button sizes and keyboard navigation were also considered for accessibility.

Engineering behind the scenes

Much of a product’s quality lives in places users never see. Here are some of the technical decisions that make Honest Lens dependable:

Several AI models, no single dependency

Honest Lens does not depend on one AI provider. Several models from different providers work behind the system. If the first model is slow to respond, a second one starts in parallel and the first valid answer is used. If one service becomes unavailable, the system moves on to the next without interruption. This architecture was especially important for users in countries under sanctions, where some services may be restricted.

Speed through smart caching

Popular stories are often checked by many people. The result of each analysis is stored for a while, so if the same text is submitted again, the report appears instantly, with no waiting and no repeated cost.

Security from day one

A tool that opens links supplied by users can become an entry point for attackers if it is not designed carefully. We implemented several layers of protection:

• Protection against using links to reach internal networks and servers. Every link is checked before it is opened, and again at every redirect.
• Per-user request limits, so the tool cannot be abused by bots or excessive use.
• Strict browser security headers to block common web attacks.
• All keys and sensitive data kept on the server side only, and a full project audit to confirm that no sensitive information is exposed in public code.
• Database access only through controlled server routes.
• Dependency checks, with no known vulnerabilities in the latest audit.

Available where the users are

The platform’s core audiences are in Russia and Iran, where some common internet infrastructure is slow or restricted. We chose hosting infrastructure accordingly, and in accessibility tests from Moscow, Saint Petersburg, Tehran and Shiraz, the site loaded without problems.

Being found: SEO and optimization for AI answer engines

A good tool helps no one if nobody finds it. Today people do not only search on Google; they also ask AI assistants. So Honest Lens was optimized both for search engines (SEO) and for AI answer engines (GEO):

• Content ready for crawlers: the home page, the Distortion Tracker and the Self-Audit are prerendered in all three languages, so search and AI crawlers see the full content without running any code.
• A specialist blog: 20 educational articles on media literacy, fact-checking and recognizing misinformation, in all three languages: 60 pages of original content in total.
• Structured data: information about the organization, website, application, articles and FAQ, in the standard vocabulary search engines understand.
• Linked language versions: every page points to its versions in the other languages, so users in each country see their own language in search results.
• Rich previews in messengers: when a link to the site is shared on Telegram, WhatsApp or social networks, it shows a title, description and image specific to that page and language.
• A guide for AI assistants: a standard file that explains to AI assistants what the site is and which pages matter most.

Principles we kept while building

For a tool about truth, how it is built matters as much as the product itself. We kept a few principles from start to finish:

• No fabricated data. No statistics, user testimonials or badges were invented. Every number on the site comes from a real run.
• No unearned endorsements. We did not use the logos of media outlets or organizations that have not worked with the project.
• Openness about limitations. The FAQ states plainly that the tool can be wrong and that users should read the explanation and sources, not just the verdict.
• Respect for privacy. No sign-up and no advertising trackers. Users are also told clearly that the text they submit is stored so that the share link works.

Results at a glance

What we learned

Honesty is a product feature, not just a value. The moments when Honest Lens says “I don’t know” or publishes its own mistakes are not weaknesses of the product; they are the main reason to trust it. In a market full of exaggerated claims, admitting your limits is itself a competitive advantage.

AI alone is not a product. Everyone has access to language models. The real value lies in the layers built around them: splitting text into claims, actually verifying links, showing the process live, measuring accuracy, and designing an experience that earns the user’s trust.

Waiting time can become part of the experience. Deep analysis takes time. Instead of hiding it, we turned it into a window onto the system’s work, and that transparency increased users’ trust.

Real multilingual support is designed from day one. Adding Persian or Russian to a product built for English always stays half-finished. Because Honest Lens was designed as a trilingual product from the start, all three versions offer the same complete experience.

Conclusion

Honest Lens shows that AI can be used to build a tool that reduces confusion instead of adding to it, as long as honesty, transparency and quality run through every layer of the product: from the “I don’t know” philosophy to verifying every link, from flawless right-to-left design to security and search visibility. We designed and built it ourselves, from the first idea to launch, with every detail in between.

You can try it right now: paste a forwarded message or a news headline into the fact-checker and see what lies behind it.