Skip to main content

Third Opinion

See something worrying? Test your concerns by receiving independent expert opinions.

Explore the process

THE CHALLENGE

Talking to current and former insiders at the frontier of AI, we repeatedly came across a question: “If I see something in my work that seems worrying or wrong to me, how can I get an expert opinion on my concern – without revealing any confidential information (yet) or making a disclosure?”

Our service 'Third Opinion' exists to solve this problem.

THE SOLUTION

We help concerned individuals working at the frontier of AI with expert-opinions on their questions.

Submit a question via our Tor-based tool to us. We then custom-select independent experts with you, handle the outreach, and orchestrate an anonymous Q&A with you. Covering a broad range of topics (scroll to “Topics we cover”).

Anonymously, securely, without sharing your confidential information.

Without pushing you towards any sort of disclosure – solely to help you judge whether what you’re seeing is cause for concern.

CONCEPT

Preparation

1. You notice a concern during your work that you think may indicate development or deployment of AI in ways that could stand in contrast to the public interest.

2. You formulate a question that would help you identify whether your concern is well-founded – without including any confidential information or information that could lead back to you or the organization you work for.

Outreach

3. You anonymously submit that question to us via the secure form (a OSS, regularly pen-tested GlobaLeaks instance).

4. You receive credentials to log back in to view your account your. As we do not store emails, we cannot send you notifications: You will have to check back in regularly.

Expert selection & approach

5. We get in touch with you via the tool to refine your question and identify relevant independent experts.

6. We approach experts to confidentially provide opinions. Experts only receive your questions if they agree to our terms & conditions around confidentiality.

You receive answers

7. We share the experts’ opinions on your questions to you via the anonymous tool.

8. We relay follow-up questions and responses, acting as a Q&A intermediary.

You decide what's next

9. Once you are satisfied with the answers, the process either ends or we can either connect you with a support organization (where you can share confidential information covered under legal privilege).

Discover the best whistleblower support organizations here. 

10. All conversation data gets deleted after 21 days of inactivity. We will never act on any of your questions without your explicit consent.

EXAMPLE

Preparation

1. Tom works at Omega AI Labs and has witnessed that the evals that are run on their models are not published.

2. He feels like this may be inadequate for a company of his scale, but he is not sure if he is overreacting. Tom is keen not to share with the outside world that Omega AI Labs [intends to] follow this practice.

Outreach

3. He downloads Brave Browser on his private machine, enters Tor, and reaches out to Third Opinion, asking “Should my lab publish their model evals?”

4. He writes down the login credentials and checks back 2 days later.

Expert selection & approach

5. The Third Opinion team responds and offers to set up an anonymous, secure call with Tom.

In the call, Tom and the Third Opinion team refine the question to make sure that experts can provide a more nuanced answer along ‘thresholds’: “What are the critical decision dimensions based on which labs should determine whether to publish their capability evals? Under which circumstances should they (not) publish their evals? What public precedent from leading labs exists for publishing of evals?”

From his research, Tom knows about two independent researchers working on this topic whose opinion he’d like to hear. In addition, Third Opinion knows about another researcher with expertise in the topic.

6. Third Opinion takes the question and gathers answers from the researchers over the coming 3 weeks. The team shares these answers with Tom via the secure communication channel.

You receive answers

7. After agreeing to Third Opinion’s T&C, the experts receive the question and, a few days later, share their answers with us. We then share these answer with Tom via the GlobaLeaks tool.

8. The Third Opinion team coordinates the Q&A process with the experts.

You decide what's next

9. Tom can reflect on the answers and, knowing the circumstances of his situation, apply the guidance to convince his team to make appropriate changes to their policy.

10. His question gets deleted from our GlobaLeaks instance after 21 days.

QUESTION POLICY

Using Third Opinion means getting help without having to reveal sensitive information.
  1. Do not include confidential or personal information in your question.
  2. Frame the question in a way that ensures your identity and of your organization are not revealed. Imagine the question being seen by a knowledgeable team member – would they be able to determine the source?
  3. Operating within this frame, make the question as specific to your concern as possible, so we can find tailored responses rather than generic answers (or none).
  4. If we believe that a question may not yield useful outcomes, we’ll work with you to refine it.

TOPICS WE COVER

We identify relevant experts together with you. That means we can cover a very broad range of topics.

This is especially the case in technical domains, where you likely know best which independent experts are most suited to answering your question. If you don’t feel comfortable approaching them directly – that is where we can help.

Before you submit your question, however please apply the following checklist to your question. You should be answering ‘yes’ to all below:

  1. Does your question relate to the development or deployment of frontier AI models?
  2. Is there a concern underlying your question relating to your organization behaving potentially adverse to the public interest (now or in the future)?
  3. Is the concern underlying your question not yet public?
  4. Are you worried about reaching out to relevant experts directly?
  5. Note: If you are uncertain about whether your question fits our scope – submit it. You can stop the process at any time and we’ll let you know if the question falls outside our purview. We will delete all data relating to your question upon your request or latest after 21 days of inactivity.
Example misconduct categories:

A concern could be as ‘small’ as “red teaming concerns are not being followed up on” to as ‘large’ as “we are making massive progress in automating ML research without any consideration for informing the public or implementing oversight mechanisms internally”.

More detailed examples:

Technical safety & development

A. Model Behavior & Control

  • Interpretability concerns
  • Unintended model capabilities
  • Control/shutdown mechanisms
  • Testing protocols adequacy
  • Capability jumps
  • Model deception indicators
  • Alignment testing practices
  • Fine-tuning safety

B. Deployment & Monitoring

  • Post-deployment monitoring adequacy
  • Safety metrics
  • Incident response procedures
  • Warning signs in deployment
  • Testing comprehensiveness
  • Production safeguards
  • Emergency shutdown procedures
  • Scaling safety measures

C. Security & Access

  • Access control practices
  • Model access protocols
  • Data security concerns
  • Training infrastructure security
  • API safety measures
  • Compute allocation oversight
  • Security testing adequacy
Organizational ethics & governance

A. Public Communication

  • Capability level transparency
  • Risk communication practices
  • Public statement accuracy
  • Marketing claims validity
  • Progress reporting accuracy
  • Safety claim verification
  • Public commitment adherence

B. Internal Governance

  • Decision-making processes
  • Safety oversight effectiveness
  • Risk assessment procedures
  • Internal review adequacy
  • Power concentration issues
  • Oversight board effectiveness
  • Stakeholder representation

C. Employee Treatment

  • Non-compete clauses
  • Non-disparagement agreements
  • Whistleblower protections
  • Internal criticism handling
  • Safety concern responses
  • Compensation fairness
  • Power dynamics
  • Retaliation patterns
Research & development practices

A. Research Ethics

  • Experiment design ethics
  • Testing boundaries
  • Research priority setting
  • Safety/speed trade-offs
  • Dual-use concerns
  • Research restriction adequacy
  • Publication practices

B. Development Practices

  • Development speed concerns
  • Safety-performance balance
  • Testing thoroughness
  • Documentation practices
  • Code review processes
  • Safety feature implementation
  • Technical debt handling
External impact & responsibility

A. Societal Impact

  • Deployment impact assessment
  • Societal risk evaluation
  • Bias and fairness issues
  • Access equity
  • Environmental impact
  • Social responsibility measures
  • Community engagement

B. Industry Relations

  • Competition practices
  • Industry standard adherence
  • Collaboration ethics
  • Information sharing practices
  • Market manipulation concerns
  • Partnership ethics
Compliance & legal

A. Regulatory Compliance

  • Compliance program adequacy
  • Regulatory requirement handling
  • Legal boundary assessment
  • Reporting compliance
  • Documentation adequacy

B. Financial Practices

  • Resource allocation ethics
  • Financial reporting accuracy
  • Investment source ethics
  • Funding use transparency
  • Accounting practices
  • Grant compliance
Culture & values

A. Safety Culture

  • Safety prioritization
  • Risk tolerance levels
  • Safety feedback handling
  • Incident learning culture
  • Pressure vs. safety balance
  • Speed vs. caution balance

B. Ethical Culture

  • Ethics training adequacy
  • Value alignment practices
  • Ethical consideration integration
  • Moral compass indicators
  • Cultural red flags
Political influence & policy engagement

A. Lobbying Activities

  • Disclosure of lobbying efforts
  • Lobbying priority alignment with public statements
  • Hidden influence campaigns
  • Industry coalition participation
  • Funding of policy research
  • Political donation patterns
  • Revolving door practices
  • Usage of third-party lobbyists
  • Political Influence & Policy Engagement A. Lobbying Activities
  • Disclosure of lobbying efforts
  • Lobbying priority alignment with public statements
  • Hidden influence campaigns
  • Industry coalition participation
  • Funding of policy research
  • Political donation patterns
  • Revolving door practices
  • Usage of third-party lobbyists

B. Policy Manipulation

  • Attempts to shape regulation
  • Messaging to policymakers
  • Accuracy of policy briefings
  • Technical information sharing
  • Risk communication to officials
  • Regulatory capture attempts
  • Policy forum participation
  • Legislative language insertion

C. Political Relationships

  • Government contract ethics
  • Political favor trading
  • Influence network building
  • Advisory role conflicts
  • Political access trading
  • Promise-making to officials
  • Backdoor negotiations
  • Relationship disclosure practices

D. Public Policy Engagement

  • Think tank funding
  • Academic influence efforts
  • Policy research manipulation
  • Public narrative shaping
  • Media relationship ethics
  • Expert voice cultivation
  • Opposition research practices
  • Grassroots manipulation

E. International Influence

  • Cross-border influence attempts
  • International regulatory arbitrage
  • Jurisdiction shopping
  • Global policy manipulation
  • International alliance building
  • Foreign government relationships
  • Cross-border compliance
  • International standards influence

F. Democratic Process Impact

  • Election influence attempts
  • Democratic process manipulation
  • Public opinion manipulation
  • Social media influence campaigns
  • Information operation involvement
  • Political targeting practices
  • Voter influence attempts
  • Democratic institution engagement
Novel concerns
  • New capability implications
  • Unexpected behaviors
  • Emergency protocols
  • Novel risk patterns
  • Unprecedented situations
Future planning
  • Long-term safety planning
  • Capability scaling plans
  • Risk mitigation strategies
  • Future governance plans

GUIDANCE ON PHRASING YOUR QUESTION

Threshold Definition
  • What [control mechanism] would be appropriate in [domain] given [dimension driving need for control 1], [dimension driving need for control 2]? Which other dimensions would drive what an appropriate [control mechanism] would be?
  • What patterns in [domain] would indicate a problem exists? What signs would warrant immediate escalation?
  • Example: “What form of a non-disparagement agreement would be concerning/ not market-standard/ potentially illegal? Especially in the context of an exit agreement and its implications of vested equity?”
Benchmarking
  • What precedent from leading AI labs exists for [control mechanism]? Note: Answers will only be based on public information.
  • What would be considered standard in [domain]? What would be considered (far) below standard?
Evaluation
  • How worrying would you consider [behaviour] be given [context]?
  • Under which circumstances would you consider this behaviour to be more or less worrying?
Specify what is (not) interesting
  • Disregard the following categories/aspects: […]
  • Please focus on the following aspects: […]

BEST PRACTICES ON SUBMITTING YOUR QUESTION

  1. Do not use a work related machine to submit a question. If your private machine is secure and trusted, you can use your private machine. A good way of making sure that your submissions are safe is booting into an ephemeral and trusted incognito OS like tail https://tails.net/.
  2. If possible, use a public network (e.g. a cafe) that you don’t regularly access.
  3. Use a locally hosted, open-source LLM to refine your question and remove your writing style. Here is a guide on how to do so: https://www.freecodecamp.org/news/how-to-run-open-source-llms-locally-using-ollama/

*With the Tor Browser or the Brave Private Window with Tor it becomes trivial to access any onionsite.

FOR ASSISTANCE SEEKERS

General
Why work with Third Opinion?

In the past, if you had a concern, you’d have to reach out to your close network, an expert directly, or a journalist, risking leaking confidential information. That’s why we created Third Opinion:

  1. Get highly qualified opinions on your questions/concerns from independent researchers at the top of their respective field
  2. No confidential info sharing: Third Opinion only shares your questions with experts, appropriately framed to ensure that you and your organisation cannot be identified.
  3. Stay in control: You decide whether or not to stay anonymous and what happens with your question and answers. Third Opinion will not act without your consent. All experts are committed to confidentiality.
  4. Help shape independent research: Help independent safety researchers by pointing them towards which questions practitioners are concerned about today
Who is Third Opinion built for?

Individuals at the frontier who reach out to us directly:

  1. You work on the frontier of AI – in some way, shape, or form
  2. You witness behaviour you find concerning – but you don’t know how to safely get a third party opinion to help you evaluate if your concerns are well-founded/ relevant
  3. You want to ensure anonymity and not disclose any confidential information to a third party

Organisations working with individuals that require peer review:

  1. Attorneys
  2. Journalists
Sharing information
Where can I submit a question?

Here: Submit a Question. Open this link using Tor browser.

What happens after I submit a question via your form?

You receive a code to log back in again. Please do so every 1-2 days, as we do not store any information that would allow us to send you notifications for security/privacy reasons.

We will then reach out to you to schedule a call via an anonymous calling software. Do not switch on your camera during the call and do not provide your real name.

In this call, we will discuss and refine your question, which (type of) experts may be best suited to your question, as well as your preferences on knowing the experts’ identities. Per default, the experts’ identities will not be shared with you.

We will then reach out to relevant experts and gather answers. This can take multiple weeks, although we can do our best to accelerate the process if time requires it.

We will share the answers through the tool. Please log back in every 3-5 days to check for answers.

You can then share follow-up questions with us, we forward them, the process continues.

That’s it. See for an example.

Can I decide to stop the process at any point?

Yes, absolutely. Just log back in, message us accordingly, and we will stop all proceedings. You will not have to provide any reason.

Does this mean I have to share confidential information?

No – the opposite. We are here to help you find answers to your questions without sharing any confidential information so that you can decide for yourself if the situation that you find yourself in has crossed any concern/risk thresholds.

Do I have to share information about myself? Do I have to share which organisation I work for?

No. We only need you to confirm that you are honestly looking for help and which type of organization you work so we can ensure independence of experts.

Do my questions go straight to experts?

No. We will evaluate them first, do our best to ensure it cannot be traced back to any specific organization/ individual, and potentially refine them with you before sharing them with selected experts.

Will you share my questions with anyone without my consent?

Never. Confidentiality is the core of our offering and we will never do anything to compromise this confidentiality.

Security, Privacy
What does “secure and anonymous submission form” mean?

We use a self-hosted instance of the most up to date version of GlobaLeaks – an open source solution used by news-rooms and non-profits globally for handling information confidentially and anonymously. GlobaLeaks is regularly pen-tested and audited.

Our self-hosted instance is hosted in Germany, our disks are fully encrypted, and GlobaLeaks encrypts the data stored in the database.

Why is your tool only accessible via the Tor network?

Given the sensitive nature of your and our work, it is better to be safe than sorry. With the Tor Browser or the Brave Private Window with Tor it becomes trivial to access any onion site.

What measures do you take to make sure my submitted questions stay confidential?

We employ strict security measures on operational and systems level, this translates to

  1. We own our data and host our own GlobaLeaks instance
  2. We host in our Jurisdiction and our services are hosted on dedicated Servers
  3. We do not use shared VMs to prevent noisy neighbours and mitigate security risks that stem from vulnerabilities such as meltdown and spectre
  4. We work on secured machines using QubesOS, where untrusted and trusted work is seperated into “Qubes” (imagine separate machines), network connections are firewalled for trusted and disabled or behind TOR on untrusted Qubes. The hard drives are encrypted and all secrets follow strict rules yielding a minimum entropy of 200 bits
  5. Our deployment is highly isolated, and access to the system is extremely safeguarded to prevent unauthorized access – for more details, please see the diagram below
image

Internal operations

  1. Contact between Oaisis Members and you implies (unless specifically agreed on otherwise) communication through secure and verifiable means – our default is E-Mail using PGP encryption. Our E-Mail provider is Proton Mail.
  2. Oaisis team members will not pro-actively reach out to you, a concerned individual, and inquire about sensitive information via email. If you receive an email looking like us that does so – be vigilant and check the signature.
  3. Internally we communicate using the same standards (ownership of data/communication, best in class secure communication methods).
  4. Our cloud storage and collaboration suite is hosted and owned by us (we avoid the big cloud).
  5. Internal tools are selected using the following minimum criteria: Open Source, high security standards, best in class solution, less is more (KISS), PoLP (principle of least privilege).
Can you guarantee my anonymity and safety?

We do our best to, but 100% security is never possible. Whilst we do our best to assure your anonymity and safety, it is important to understand that a large amount of vectors is on the client side. We can only secure our side of the equation, be mindful and follow best practices of security at all times. See the question above for details on the measures we take.

Interaction with experts
Who are the ‘independent experts’?

Individuals who are not employed or in any meaningful way involved at an organization related to your organisation. This could mean an independent researcher, a researcher at a research institute, in academia, a think tank employee, lawyer, etc.. You will always have control over who your question is shared with (see question “Will I know who the experts are”).

How can you possibly have a relevant expert for my specific question readily available?

We likely don’t. We recruit them for your question. You can point us to independent experts that you think are most fit to answer your question.

Can I get a list of the independent experts?

No, we don’t publish a list to protect the privacy of our experts. Once you have a question you’d like support on, we will discuss options with you.

Will I know who the experts are?

The identity of the expert can be shared with you if the expert agrees and you want to know their identity. You can also decide that you don’t want to have your question shared with any experts you do not know the identity of.

Why would experts be interested in my question? Why would they work with you?

Because they are like you and me: They care about responsible development of AI. They also want to make sure that their research matters, reaches those that need it, and that they’re considering the most relevant considerations in their own research.

How do you ensure that experts keep my question confidential?

We employ a two-pronged process: We first ensure together that the phrasing of the question does not contain any confidential information or information that would identify you or your organisation.

In addition, our terms and conditions ensure that they won’t share your question, the Q&A, or the fact that they received it via Third Opinion with any Third Party. Our experts have to accept these T&C before accepting your question.

What they are allowed to do is have your question inform their research and their publications. Any other commitment would also not be feasible, as they will have very likely been working on your or a tangential question before our outreach – otherwise they would not be experts.

Do you guarantee that one or more experts can be found for my question?

No. We will do our best but cannot guarantee it.

You are free to point us towards specific experts that would be well suited to answer your question.

Do you guarantee the correctness of the expert’s assessment?

No. We do our best to bring you the most qualified opinions, but our experts are human.

Can experts use my questions for their own research?

Yes. They are however not allowed to disclose that their inspiration was a conversation held via Third Opinion.

About Third Opinion
Is Third Opinion associated with any AI lab, government, or any other organisation that would cause you to have a conflict of interest?

No. We are a non-profit and are offering this service because we care about supporting individuals that want to make this world a better place.

What’s the relationship of AIWI (formerly OAISIS), Third Opinion, and the Whistleblower Netzwerk e.V.?

Third Opinion is one of multiple projects under the umbrella of the AIWI (formerly OAISIS) brand. AIWI is a project currently hosted under Whistleblower Netzwerk e.V., Germany’s largest and oldest non-profit organization aimed at supporting anyone who wants to speak up about misconduct in industry or government. Note that this does not mean Third Opinion provides Whistleblowing Support services. We are only here to help you understand whether what you are seeing is cause for concern.

Are you ‘enemies’ of organisations working at the frontier of AI?

No. We support organisations and individuals working in those organisations that lead the way in developing AI in the public benefit. We support those organizations and individuals who act honestly, responsibly, and transparently. We acknowledge that current incentive systems are set up in a way that only exceptionally thoughtful, strategic, and capable leaders can resist the short-term benefits offered by ‘skipping’ on safety and fostering cultures focused on in-group loyalty. In addition, we share an interest with labs in preventing ‘boy who cried wolf’/’false positive’ situations, in which concerned individuals might leak reports with the public that turn out to not in fact be relevant issues.

FOR EXPERTS

Will my identity be published/ shared with anyone?

You are in control of whether your identity is revealed to the Assistance Seeker. We will not publish your identity or your involvement with Third Opinion without your consent.

Can I use the questions to inform my own research and publications?

Generally Yes, as long as there is no risk of this allowing a third party to identify that you had a conversation with someone via Third Opinion or that a conversation with someone from an external organization inspired your research: See our Terms of Service: Allgemeine Geschäftsbedingungen (AGB).

Are there any answers I should not give?

You should never advise Assistance Seekers to break any contractual obligations.