Answer

Yes, IP protection would apply. This is dependent on the countries involved and their national law and is something ISPs have battled for many years on whether they are liable for users acts. Generally, courts have found users and not service providers liable but this is being changed in some countries to make service providers more responsible/police their users. Replication would still infringe some's IP even if AI was used. It is likely to be the person giving the commands to be found to infringe. Similarly, overcoming protection measures is a form of copyright infringement in many countries.

Answered by Anna Orchard

Answer

Use commercial versions of the tools and instruct them not to store data or prompts or use them in future training. This should keep what you are inputting confidential (worth checking terms and conditions of whatever LLM you use but generally commercial contract includes duty to keep inputs confidential). You can protect concepts in the same way as if you had come up with the idea without using LLMs - copyright will happen automatically, you can apply for patents and trade marks etc. If you think there is value in registering for the business, get professional advice as using an LLM hasnt changed the rest of the process.

Answered by Anna Orchard

Answer

Yes, we are limiting functionality. That is the point of guardrails. The point of your question may be that guardrails also limit desired capabilities. This may be true to some extent, but it appears to be a relatively small effect.

Answered by Maury Shenk

Answer

Both core capabilities and guardrails deserve attention. Consider a hammer -- it appears impossible to prevent it from being used as a weapon by changing its 'core capabilities'. So we need to depend upon restraints on its use. Controls on multi-agent systems are indeed challenging, but I am confident that these will improve over time, subject to the challenge that agent capabilities will be improving at the same time.

Answered by Maury Shenk

Answer

It depends upon the situation. For some types of problems (e.g. does this LLM answer talk about a prohibited subject?), guardrails can get to or near 100% compliance. In other cases, guardrails may only be limiting the probability of failure. But that is a significant contribution if the reduction in probability is large, and may significantly reduce risk. In many societal domain, quantum of risk is a key consideration.

Answered by Maury Shenk

Answer

This is a very complex problem to which the market does not yet have an answer. There are some partial answers like blocking access to particular tools, or blocking Internet access completely (or severely limiting it). Also, the response depends upon the specific attack.

Answered by Maury Shenk

Answer

The future is hard to predict! I expect that current guardrail methods will evolve steadily, and new approaches will emerge. Methods may also generalize. For example, providing additional context is likely to be a durable guardrail, but embeddings-driven RAG may play a lesser role over time in providing context.

Answered by Maury Shenk

Answer

Various options are possible. The main ones are (1) modifying the output (regeneration is one way to do this) and (2) refusing to answer (the main approach of leading models).

Answered by Maury Shenk

Answer

Pacing the frontier and controlling models are two different solutions to AI risks. Both may have a role.

Answered by Maury Shenk

Answer

I do not believe that it is possible to effectively tell an AI model not to hallucinate. Hallucination controls need to be external to the model -- see this Satya Nadella post: https://x.com/satyanadella/status/2108931348857827686. Such external controls are what we are building at Ordinary Wisdom.

Answered by Maury Shenk

Answer

While Japan is more permissive, there are exceptions to what is permitted and its not a clear "no risk" situation. https://www.ibanet.org/Japans-AI-training-exception Very few countries have identical IP laws and approaches and even when they do, government policy and/or commercial attitudes and services tend to force them to come closer together over time. I'd expect the stricter approaches to win on this over time.

Answered by Anna Orchard

Answer

China can be a challenging market, both to companies from outside wanting to enter their market and also from products and services originating from China. Having IP helps. You can obtain IP in China to protect your activities there and also in other countries to protect imports from China. Its not a new challenge, but it is one where businesses have successfully used IP to defend what is theirs for many years and AI will not change that.

Answered by Anna Orchard

Answer

Vibecoding gets you to a product/service/market quicker than historic routes but still has the same risks and opportunities. I'd recommend you ignore how you got to the product/service at it doesn't matter to IP. Look at what differentiates you from competition, what someone else might want to copy and work out if the business can support the cost of IP registration and if it would benefit from the IP. If so, get professional advice. If there are big or aggressive businesses in the field, you need to look at what they are doing/think is theirs as those tend to be the biggest risks to a new entrant.

Answered by Anna Orchard

Answer

You need to look at the terms and conditions of the system you are using. This is also relevant on whether they can take the data you input and use it for future training etc. Generally, if you are the originator/creator of a work that has IP you are the first owner of the IP rights. However, by inputting it into a system you might (depending on the terms and conditions) be giving other parties rights to use your IP. You might also be disclosing your concept (bad if you are thinking of applying for a patent).

Answered by Anna Orchard

Answer

Trade secrets are also referred to as "knowhow". They are protected through confidentiality/secrecy but don't give you any ability to stop someone who independently has the same knowledge/idea. Trade secrets are valuable and are grouped under "IP" when you give someone a license etc but they aren't formally protected, just useful information you might have documented or in your head. It doesn't really relate to software as any software will also have copyright, although the knowhow on configuring a server or database for the software might fall in this category.

Answered by Anna Orchard

Answer

My understanding is that GitHub allows a change in license type/conditions for future branches and new downloads but that you cannot retroactively change the license (so anyone who has downloaded the code or who creates a new branch or project themselves would be able to do so under the MIT license).

Answered by Anna Orchard