Sustainable AI anti-fakenews system

Port our existing ML models for fakenews detection to SingularityNET and use them with user-facing engagement frontend and DAO on Cardano.

Experienced ML team led by a senior AI researcher&executive. The off-chain ML backend is already running. Team: https://www.verifee.ai/en/

We are the team behind the emerging AI-based anti-fakenews platform Verifee.ai.

Main challenges for Verifee that we're addressing by the proposed project:

1. How to make it transparent, trustworthy and resistant to manipulations. -> We'll make it decentralized by deploying it on SingularityNET.

1. How to fund, govern and maintain it in the long run. -> We'll create decentralized Cardano-based funding mechanism that will in future lead to DAO.

1. How to incentivize the newsreaders to use it, even though it might introduce them to some "discomfort" by interfering with their comfortable ideological rabbit holes. -> We'll integrate the Verifee frontend with a Cardano wallet and introduce a "play-to-earn" strategy. 

What important societal problem is Verifee addressing:

The current rate of production and large-scale consumption of fake news leads to dangerous ideological polarization in society. It fuels growing tensions and hatred among increasingly mentally isolated groups of people. It can lead to delusional pathological states that psychiatrists call "folie à millions", i.e., shared psychosis. This can lead to an amplified radicalization of such individuals and can be ultimately devastating for our whole society.

How is Verifee addressing this problem:

Verifee is NOT a fact-checker. It is NOT supposed to tell what "the truth is". Instead, it infers the credibility of news articles by identifying various manipulative techniques (explicit and subliminal) that are attacking the reader's cognition in order to pull her towards some simplifying ideology. Verifee user frontend is a web browser plugin that exposes these manipulative attacks to the reader. It also patiently engages the reader in a continuous learning process of how to identify such manipulation on her own. Verifee backend is based on an ensemble of machine learning (ML) models, mostly from the domain of Natural Language Processing (NLP), such as pre-trained language models.

Why we're applying for Fund7:

We want to bring the platform to the decentralized world so that it eventually cannot be manipulated or shut down by any single entity (politicians, corporations, ideological groups, etc.). We want to utilize Cardano to create a sustainable decentralized funding model for the system's long-term development and operations, eventually leading to a full DAO governance model. This funding model shall also incentivize the users by rewarding them for their interaction with the Verifee fronted while reading the news (i.e., introducing a subtle "play-to-earn" dynamics).

Deliverables of this Fund7 project are:

1. Decentralized ML backend – port all Verifee models to SingularityNET and make them a functional part of the SingularityNET decentralized marketplace.

1. Verifee frontend integration with Cardano as the AGIX wallet – develop a user-facing AGIX-reward incentivization mechanism.

1. Prototype of a decentralized funding mechanism – a combination of SingularityNET ML services earnings, user AGIX donations, and dedicated SPO.

 The overall high-level architecture scheme is in the attached PDF.

Ad 1) Decentralized ML backend

Verifee already has a functioning MVP consisting of centralized ML models, API, and a minimalistic web frontend. It already has some public non-gov and non-profit traction, see https://www.verifee.ai/napsali-o-nas-5/

We will port all the necessary Verifee ML modules to SingularityNET and make them available on the decentralized marketplace. The rest of the system will then query them on-demand via SingularityNET and pay for these queries by AGIX. This will allow for anyone to query our services as well (they can be suitable in a wide range of projects, primarily NLP), hence generating AGIX earnings.

Initially, this part of the proposed solution will be handled on Ethereum. However, as SingularityNET moves fully to Cardano, so will the Verifee ML services.

The ML models themselves will initially run on standard centralized virtual machines with deployed SingularityNET sidecar demons. Later on, these will be moved to a decentralized NuNet environment (once NuNet is ready for such a use-case). This will finalize the gradual decentralization transition.

The current version of Verifee has been developed for the Czech language. As a part of this project, we will extend the functionality to cover English too. We use state-of-the-art ML/NLP methods which are language-independent. Some manual English training data annotations will be needed. We would like to prepare such an ML architecture that will be extensible to any further language environment by the developer community in the future DAO.

Ad 2) Verifee-Cardano user frontend

Verifee frontend consists primarily of a browser plugin accompanied by a web application for detailed analyses or interactive tasks. It is being developed outside of this Fund7 project (the development is self-funded, or will have different sources of funding, either CSR, or a university R&D grant).

The primary functionality of the frontend is to deliver findings about each news article to the user in a psychologically optimal way. The results for each article will be obtained by the backend querying our ML services on SingularityNET.

The proposed Fund7 project shall integrate this frontend with a Cardano wallet with a specific AGIX token functionality.

The user will be rewarded with AGIX token for her proactive interaction with the Verifee plugin on a daily basis. This will drive the incentivization of those users who would otherwise never have used such a tool, as gaining such a token will be presented in a manner linked to a familiar currency and thus associated with the general concept of earning money/rewards.

To provide an illustrative example, let us imagine the following scenario:

- The user opens an anti-covid fake news article; the Verifee plugin becomes bright red, indicating that an outright manipulation is detected.

- The user clicks the plugin and gets the prompt: "It seems like this article has some hate speech in it. Do you think you could find it and show it to me?"

- If the user selects a portion of the text that at least very roughly corresponds to what the Verifee ML backend detected as hate speech, she receives e.g. 1 AGIX reward (this has some capping, e.g. once a day).

This project shall deliver a prototype of such a Cardano-integrated interface with basic functionality.

Future versions shall have much broader engagement and incentivization functions, better cheating detection, etc. They shall allow for user AGIX donations to the system and other operations with the AGIX token (such as cashing it out or paying for some future consumer-oriented AI services on the SingularityNET marketplace). Social aspect in the form of tailored rankings among user's close friends and family may also be employed in the app to establish both motivation and a sense of peer pressure primarily when onboarding.

The specific dynamics of psychologically efficient user interaction is developed in cooperation with volunteering researchers from Czech National Institute of Mental Health (group of prof. Jiri Horacek).

Ad 3) Decentralized funding mechanism

It is necessary to create a mechanism that can fund the following in the long run:

a) backend ML services on SingularityNET marketplace

b) compute resources (eventually on NuNet)

c) decentralized community DevOps/MLOps team (including an ML/NLP researcher/engineer) and future DAO governance

d) rewards for user incentivization

The project will create a prototype of a Cardano-based mechanism and fund pool (Verifee Fund) that will transparently accumulate ADA/AGIX from the following sources:

- earnings by our ML models deployed on SingularityNET providing paid NLP services to third parties

- earnings of a dedicated Verifee Cardano Stake Pool

- ADA/AGIX donations of users who wish to support the cause

The mechanism will then redistribute the acquired resources according to the respective needs of (a)-(d).

It is clear that at this moment this will not create sustainable financing immediately – mostly because of low traction of the SingularityNET marketplace (so no real income is to be expected very soon from the ML services) and because of an initially very low number of users willing to donate. Therefore, the main source of initial funding shall be the SPO, until other sources gradually kick in.

Future streams of financing can be extended by further sources such as prediction market monetization and/or NLP data labeling based on crowdsourcing approach from users' interactions with the frontend.

Budget allocation

We are going to be the 5-member team:

- 3x ML/SW engineer (we already have these roles)

- 1x blockchain engineer/developer (we need to hire this role)

- 1x senior ML researcher and project lead (we already have this role)

The monthly allocation per person is 2200 USD; the time allocation of each person will be based the average market income benchmark of that respective role in Prague, CZ. E.g., if the average benchmark income of the role is 3000 USD/month, then the committed time allocation will be 2200/3000 = 73 % of Full-Time Equivalent.

The duration of the project is planned to be 10 months.

The total budget: 2200 * 5 * 10 = 110 000 USD

Why 10 months of non-fulltime work on this project instead of fewer months of fulltime work with the same budget?

We considered the latter as the option, too. Nonetheless, we feel that the SingularityNET ecosystem is still quite immature with regards to Cardano, and so are other smart contract dev tools for Cardano. Thus, we eventually decided that we would rather like to work on the decentralized Verifee a little slower while leaving more time for the rest of the Cardano ecosystem to grow, mature, and make our work more efficient.

External Funding

The team must work also on other parts/features of the Verifee system that are beyond the scope of this Fund7 project. For example, the frontend application itself (especially with UX and engagement dynamics, which are super important), R&D of new ML models (e.g, semantic, pragmatic and graph feature extractors), training data acquisition and annotation, general fake-news research, integration with partners, etc.

It means that the team will work full-time on Verifee anyway, optimally distributing its capacities between the proposed Fund7 project and the other activities funded externally.

The external funding has so far been mostly from private sources and CSR grants. We're now actively applying for university R&D grants and philanthropic grants provided by individuals. If we succeed, we'll be able to grow our team beyond 5 people and reach the milestones faster.

Milestones and Definition of Success

3 Months

- all Verifee ML models deployed as services on SingularityNET

- Verifee Cardano Stake Pool up and running

6 Months

- Verifee backend and whole query pipeline works with SingularityNET

- prototype of Verifee frontend integrated with Cardano wallet

- Verifee ADA/AGIX Fund (smart contract) deployed, it collects AGIX earnings from Verifee SingularityNET services and ADA from Stake Pool rewards

- Verifee Cardano Stake Pool has enough delegators so that it can generate rewards to Verifee Fund

10 Months (planned project duration)

- Minimum Viable Product of Cardano-integrated Verifee frontend app (plugin)

- fronted has "play-to-earn" incentivization functionality with AGIX reward transfer to individual users Verifee wallets based on their engagement with Verifee

- automated payments from Verifee Fund to ML services on SingularityNET

- Verifee works for English too

12 Months

- international outreach and dissemination through the Cardano community

- dissemination and user onboarding in Czechia through O2 Foundation (Nadace O2 - https://nadaceo2.cz)) who already gave us commitment for the support, and also dissemination to Czech schools (teachers and students) through the CSR programme O2 Smart School (O2 Chytrá škola - https://o2chytraskola.cz))

- first users onboarded (in beta testing programme)

- target at 12 months: 10k users worldwide, 10 % of Czech schools testing Verifee