Detectors and generators compete in a dynamic loop: detectors sharpen their ability to spot synthetic media, while generators push to create more convincing fakes.
SN 34. Rank 30 of 32 by emission
BitMind
Emission1.0%of the block reward · rank 30 of 32
Participants117 miners · 5 validators
Registration0.5 TAO
Verified2026-09-09links checked and scores merged
The programmedia · github.com@45e92d8 · read 2026-09-19
The work
A media-provenance classifier: a model that takes an image, video or audio clip and returns class scores for real against synthetic. Detector miners submit the model to the platform and run no servers. Generator miners make synthetic media on demand for validators. What is paid for is, per lane, the reigning best classifier on the current benchmark, and for generators verified output that fools the detectors.
own docs · raw.githubusercontent.com
Detectors and generators compete in a dynamic loop: detectors sharpen their ability to spot synthetic media, while generators push to create more convincing fakes.
How it is scored
Per modality a detector is benchmarked on public, private holdout and generator-made samples and scored by sn34_score, which combines classification performance, the Matthews correlation coefficient, with probability calibration, the Brier score, under round-specific sample weights. A challenger takes a lane's crown when its sn34_score is at least 0.01 above the sitting king on the same benchmark version. Generators score pass rate x a volume ramp x a model-price multiplier.
own docs · raw.githubusercontent.com
combines classification performance (MCC) with probability calibration (Brier score)
How the pay splits
Fixed lanes: image 0.40, video 0.40, audio 0.04, generators 0.16. Inside each detector lane the current king takes 0.85, the previous distinct king 0.10 and the one before 0.05; unused slots roll up to the king and an unresolvable king's share burns. The generator pot is proportional to score among qualified miners, image weighted 0.30 and video 0.70. All detector lanes burn while GAS reports emissions disabled.
code · raw.githubusercontent.com · gas/koth_weights.py
KOTH_LANE_RESIDUAL = (0.85, 0.10, 0.05)
Who judges
Five validators were active on the 2026-09-09 chain snapshot, of 64 permitted, against 117 miners. Weights are committed then revealed, at most once per 100 blocks, with weights_version 404020. A validator generates synthetic media, challenges generator miners, and every epoch fetches the reigning kings per lane from the platform's kings API, caching the last answer if it is down. Detector benchmarking runs on the platform, not on the validator.
chain · hyperparams-2026-09-19.json
commit_reveal_weights_enabled True, weights_rate_limit 100, weights_version 404020; chain-2026-09-09.json active_validators 5
How often it turns
Emission settles every 360 blocks, about 72 minutes, and validators set lane weights on the reigning hotkeys every tempo. The puzzle turns per benchmark round: a throne is held until a challenger clears the margin, and when a new benchmark version ships the king is marked defending until re-evaluated. Generator qualification looks back 7 days, base rewards 24 hours. A new registration is immune for 5,000 blocks, about 16 hours.
chain · hyperparams-2026-09-19.json
tempo 360, immunity_period 5000
Who buys
Developers and enterprises. The API pricing page lists Free at $0 a month for up to 100 requests, Pro at $100 a month for 10,000 or more, and Custom for large organisations with SLAs and on-premise deployment. Consumers get a web app, mobile apps and browser extensions. No customer is named on the pages fetched and no dated usage figure appears.
own docs · docs.bitmind.ai
Pro | $100/mo | Teams and power users who need 10,000+ requests/month
What has been gamed
Overfitting the test: each round has holdout datasets unreleased until it ends, and each hotkey gets one counted submission per registration, so a miner cannot spam variants. Recycling a hotkey: a replacement hotkey cannot inherit the prior owner's generator eligibility at the same UID. Generator wash: pay only in a modality whose 7-day fool rate clears 2 percent for image, 1 for video, on at least 20 samples.
own docs · raw.githubusercontent.com
A replacement hotkey cannot inherit the previous owner's eligibility at the same UID.
3 of 7 fields rest on the chain or the code. A manifest is the site's reading of the subnet's program, not a review; each field links its source. All programs · What a field means · Correct a field
What it says
The subnet in its own words: the line the index took from its front door, and what the team has said on the record. All 15 quotes in the map.
Their subnet
A deepfake detection subnet that became GAS: generator miners forge synthetic media, detector miners submit models, validators evaluate both, and the whole loop co-evolves under aligned incentives.
Deepfake detection is one of the toughest challenges in computer vision because different generative models create various types of fake images or videos
BitMind's decentralized architecture is well-suited to this problem. By leveraging a network of diverse models and miners, we can continuously adapt and improve deepfake detection methods
The incentive mechanism we implemented has effectively rewarded accurate deepfake detection models while mitigating malicious gaming
BitMind's mission is to create the best deepfake detection solution on the market, demonstrating the potential of decentralized AI and building the infrastructure for AI engineers to develop powerful end-user applications.
We built the deepfake detection subnet on Bittensor (Subnet 34) which incentivizes identifying and mitigating AI-generated image content.
Bittensor, in their words
Bittensor is the decentralised network where developers compete for automatic rewards; the subnet exists partly to demonstrate the potential of decentralised AI. On the enterprise site the network is not named.
BitMind's mission is to create the best deepfake detection solution on the market, demonstrating the potential of decentralized AI and building the infrastructure for AI engineers to develop powerful end-user applications.
Instead of one team trying to keep up with thousands of deepfake creators, we've built BitMind on top of Bittensor, the decentralized network where hundreds of developers can come and compete to create the best detection algorithms, with automatic rewards for superior performance.
Frames competition network engine layer infrastructure co-evolution self-evolving
What it shows
Explained from zero by the index, from the subnet's own materials, for someone who has never heard of Bittensor.
SN34 BitMind sells a counterfeit-spotting desk for images, video, and audio, where strangers compete to build the best detector and other strangers are paid to forge better fakes to test it.
The commodity, explained from zero
A deepfake is a picture, clip, or voice made by an AI model and passed off as real. A bank's counterfeit desk looks at a note and says real or fake; BitMind does the same for media. The product is an API, an address a program sends an image to and gets a verdict back, at api.bitmind.ai with a key, plus an upload page, a browser extension, and phone apps.
It has a price because companies act on media they did not make; the front page puts it as: "We verify the authenticity of every data point before your AI acts on it." Pricing has three tiers: free for 100 requests a month, Pro at 100 dollars a month for 10,000 or more, and custom contracts for "Large organizations with bespoke needs, SLAs, and integrations." The docs name consumers, developers, and enterprises as paths, and finance, social media, and government as sectors. No customer is named.
Why it is on Bittensor at all
The claim is adaptation, not cost. A fixed detector goes stale as generators improve; BitMind says it wants "a dynamic, self-evolving service that instantly adapts to new threats", and the subnet is how: detectors come from an open contest, and a second set of miners is paid to make fresh fakes for them to catch. That is where the comparison breaks: a bank's forgers are on staff and their notes never leave the vault, while here the forgers are strangers paid in the subnet's token, and their fakes feed the weekly test set. Nothing is said about cost against a centralized detector.
How the work gets done
Detector miners submit a classifier, a model that takes an image, video, or audio clip and scores it real or synthetic; they do not run servers. Generator miners make synthetic media on demand and earn a base reward for valid output times a multiplier, capped at two, for how often they fool the detectors. Validators, the judges who decide pay, test detectors weekly on fresh real and generated media and score them on accuracy and on how well calibrated their confidence is, combined into one number, sn34_score. Emission, the subnet's share of new tokens, is split 40 percent image, 40 video, 4 audio, and 16 to generators; within each lane the reigning best model takes 85 percent and the two previous champions 10 and 5, and a challenger must beat the king by a 0.01 margin after a 48-hour defending window.
How you would know it works
The detector at bitmind.ai/detect accepts an image or video upload without a stated sign-in; that is the artifact. The front page claims "95% accuracy on real-world deepfakes" with no date, and its usage counters render as zero to a fetcher.
What is missing
On chain the subnet has no URL, no description, and no Discord; a wallet reader has no front door. The chain contact is intern@bitmind.ai while the docs use hello@bitmind.ai. The docs' API Reference link is broken and the full reference sits behind an account. There is no status page despite a 99.9 percent uptime claim. Registration cost and the deregistration policy are not stated.
Go deeper
Sources for this explainer
- https://bitmind.ai
- https://bitmind.ai/detect
- https://bitmind.ai/product/api
- https://docs.bitmind.ai
- https://docs.bitmind.ai/bitmind-overview.md
- https://docs.bitmind.ai/getting-started.md
- https://docs.bitmind.ai/pricing.md
- https://raw.githubusercontent.com/BitMind-AI/bitmind-subnet/main/README.md
- https://raw.githubusercontent.com/BitMind-AI/bitmind-subnet/main/docs/Incentive.md
- https://raw.githubusercontent.com/BitMind-AI/bitmind-subnet/main/docs/Discriminative-Minin
Metaphor: a counterfeit-spotting desk that pays its own forgers. Every claim is drawn from the evidence set or the subnet's own materials; "(inferred)" marks a conclusion rather than a quote. Corrections.
How legible it is
Four audiences, four questions each, rated on what a first-time reader can find in five minutes. The scale measures findability, not quality.
On-chain identity 3/7- subnet_name
- BitMind
- github_repo
- github.com
- subnet_contact
- intern@bitmind.ai
- subnet_url
- not set
- discord
- not set
- description
- not set
- additional
- not set
Stakers and validators●●●●● 3.0
Should I allocate here? · rank 18 of 32 for this audience
Q1What it is●●●●● 4
Detectors and generators compete in a dynamic loop: detectors sharpen their ability to spot synthetic media, while generators push to create more convincing fakes.
Output is detection classifiers sold as API requests priced per month, the chain has no description to match
Q2Who it is for●●●●● 2
95% accuracy on real-world deepfakes, sub-second latency, deployed across 100+ countries
undated, the front page counters for weekly detections and monthly active users render as 0+, no named customer on own pages
Q3How it resists gaming or fails●●●●● 4
Incentive.md one click from the README with the sn34_score formula, emission split image 40, video 40, audio 4, generators 16 percent, 85/10/5 king shares, a 0.01 margin and a 48 hour defending window, generators exist to challenge detectors, no failure modes stated
Q4Identity and documentation●●●●● 2
identity 3 of 7, no url, discord, description or additional on chain, github resolves, chain contact intern@bitmind.ai differs from hello@bitmind.ai on the docs, no staker page
Miners●●●●● 3.8
Can I compete, and what wins? · rank 11 of 32 for this audience
Q1What it is●●●●● 4
Miners submit media-provenance classifiers across three modalities: image, video, and audio.
input is media, output is a logit array per modality (3, 4 and 2 classes), gascli push command, sandbox rules, no worked example seen
Q2Who it is for●●●●● 3
Miners do not need to host hardware for inference.
min_compute.yml says miner compute varies by model, registration cost not stated (chain burn 0.5 TAO today), competitiveness stated as an 80 percent entrance exam then a 0.01 margin over the king
Q3How it resists gaming or fails●●●●● 4
scoring in prose with MCC and Brier normalisation exponents, lane and king shares, generator multiplier capped at 2.0, a 48 hour defending window, no worked example numbers, no deregistration policy
Q4Identity and documentation●●●●● 4
Mining, Discriminative, Generative and Installation guides linked from the README, release v5.0.0 dated 2026-09-06, repo pushed 2026-09-09, guides undated
Buyers and enterprises●●●●● 4.3
Can I use this today? · rank 6 of 32 for this audience
Q1What it is●●●●● 5
Authorization: Bearer $BITMIND_API_KEY
with base URL https://api.bitmind.ai and a curl example to /detect-image, a live detector at bitmind.ai/detect, browser extension and mobile apps
Q2Who it is for●●●●● 5
Large organizations with bespoke needs, SLAs, and integrations
for the custom tier, Free $0 for 100 requests a month, Pro $100 a month for 10,000+, customer type named per tier
Q3How it resists gaming or fails●●●●● 4
99.9 percent uptime SLA claimed on the API page with custom SLAs and 24/7 support for enterprise, contact form promises a reply within 24 hours, no status page (status.bitmind.ai does not resolve), front page uptime counter renders 0%
Q4Identity and documentation●●●●● 3
getting started page has endpoints but the API Reference nav link is broken and the full reference sits behind an account at app.bitmind.ai, hello@bitmind.ai on the docs does not match the chain contact intern@bitmind.ai
Newcomers●●●●● 3.0
What is this and why does it matter? · rank 22 of 32 for this audience
Q1What it is●●●●● 3
Welcome to the documentation for BitMind, the AI content detection platform.
plain on the site and docs, the README opens with the GAS mechanism, the chain carries no description at all
Q2Who it is for●●●●● 3
dynamic, self-evolving detection system that adapts to new threats
contrasted with fixed detectors, the difference is adaptation rather than decentralization, no comparison table
Q3How it resists gaming or fails●●●●● 4
live detector that accepts an image or video upload without a stated sign in, 95 percent accuracy claim undated, leaderboard at app.bitmind.ai is JS rendered
Q4Identity and documentation●●●●● 2
subnet_url and description are not set on chain, the site was reached through the repo homepage, subnet_name matches the site, the site has FAQ and contact pages
Provenance
How this rating came to be. Verified means the scorer fetched the page and the words are on it; secondhand means concluded from code, absence, or a third party. The link badge is a separate automated check made before publication.
scored by agent b 2026-09-09, rubric v1.0merged 2026-09-09links verified 2026-09-09: 39 live, 0 unreachable, 1 manual
Something wrong? Corrections of fact are applied as they arrive during the window; rating disputes are batched at its close. How to file one · GitHub issue · email.