Fact Checking Additions and Improvements

Here is a short list of recent notable AI developments that we think professional writers and publishers should be paying attention to:

National AI Policy. The White House released the national AI strategic framework, outlining a broad set of goals and policies that advance AI development, reduce regulatory red tape, and target infrastructure and data investment. We think this is a positive and necessary step to accelerate the most important technology of our generation, akin to 1990's era regulatory decisions that paved the way for the development of the internet. To see an interesting retrospective of that regulatory framework, plus some lessons that we might apply this time around: https://knowledge.wharton.upenn.edu/article/internet-policy-20-years-later-u-s-get-right/

GPT 5 Launch. GPT 5, the latest and most powerful model from OpenAI, launched yesterday. GPT 5 is said to be better at specific task based benchmarks, such as coding and chain of thought reasoning. We  know from experience, however, that sometimes the launch benchmarks and the reality of what our customers do in their daily jobs do not connect. Given that, we are testing GPT5 right now. If it outperforms the reasoning models we are currently using for Fact Checking (see below), we will implement it within days. 

Reasoning Models Don't Always Say What They Think. A white paper from Anthropic studied an interesting phenomena in which advanced reasoning models (such as Claude Sonnet) that describe their chain of thought reasoning often omit steps or get some of their own reasoning wrong in the description. This is why Symbolics' fact checking (which currently uses two reasoning models) provides the writer or editor an audit trail to sources to confirm the conclusion (see below). The goal is not to completely outsource fact checking to AI; it is to make humans much more efficient.

And here's what's new in Symbolic: 

Fact Checking Reasoning, Conclusions and Confirmation

As a follow on to our release of comprehensive fact checking, Symbolic has released a "v2" version which is using the most sophisticated reasoning models that are available to date. By separating out the extraction of facts and the checking of those facts, Symbolic is getting the best out of multiple models and delivering advanced reasoning for its support or refutation of the facts in a draft. We note that at times the model will refute a fact for very precise reasons (such as "1.7 is not almost 2") and that level of detail is very useful for editors or writers who can draw their own conclusions. 

As always, Symbolic will check a draft that is created on the platform or imported from outside, and uses curated sources placed on the Research canvas,  predesignated Trusted Sources, and the web. This provides the most comprehensive fact check solution to professional writers and publishers available anywhere. 

Fact Checking: Improved Source Targeting

In addition to improved reasoning, Fact Check v2 better targets the portion of the source documents that backstop the reasoning. Symbolic previews the relevant portion of the source document below the reasoning conclusion, and a click will bring up the entire document, with the relevant portions highlighted. This saves an enormous amount of time, and allows the fact checking process to be just like working down a checklist.  

Rechecking by Fact

Fact Checking now allows the user to recheck any individual fact. This might be because they would like to find different/additional sources, or because they are unsure of the conclusion and want to go deeper. If the user initiates a recheck, Symbolic will not use the same sources again. 

That's it from us this week. We are always working to improve Symbolic and make your experience more efficient and enjoyable. We would love to hear your feedback and any suggestions you may have for future updates. Don't hesitate to reach out to team@symbolic.ai.