I Spent $329.7 Just In GPT Credits For This Update

This Wraith Scribe update makes it arguably the best AI writer out there (if it wasn't already). Has the most up-to-date techniques in AI content evasion, tables and chart generation, and now writes with web-search based content. Read below to find out more.

Key Takeaways

• Up-to-date AI content evasion; passes Copyleaks consistently.
• Uses web-search from authoritative sources to write high-quality articles, without plagiarism.
• Can now generate flowcharts and tables to illustrate ideas more clearly.
• Includes other decorations like FAQ sections and key takeaway points in the intro for better SEO.

Motivation

This is a very large update and is motivated by many things:

1. A few customers brought up that copyleaks was able to detect Wraith's writing as AI-generated.
2. Bug fixes that I've been wanting to fix, but might not make sense to fix without rewriting the entirety of Wraith (which is what this update mostly is)
3. My own motivation to create much better engagement for readers
4. Minimizing factual hallucinations from GPT

Why you should care

AI Content Evasion

AI content evasion isn't that important for SEO, as is stated by current Google guidelines. The main problem is that guidelines change constantly. Thus, it's not wise to use AI content generating services to create blogs / articles if you want it to be evergreen. You might have some quick / cheap results in the short-term, but there's a very real risk of getting severely penalized when (not if), Google changes their policy and makes all grandfathered AI-generated content rank lower.

This equates to all your time / effort / money spent using other AI content generators (think ChatGPT, copy.ai, etc) have an ROI of 0. So, in order to future-proof our customer's content, we're always constantly improving our engine to be more humanlike.

Better Writing Styles / Higher Engagement

Higher engagement for readers = longer they spend on your site. This is good for SEO since Google generally do not want to recommend pages with high bounce rates. Additionally, Wraith generates content from the first-principle basis of trying to automated content really good and engaging so that your readers want to keep reading your articles intrinsically. And not just because your article's filled with some keywords.

Factual Hallucinations

GPT can say things that are false. Thus, this update tries to mitigate it by mimicking the results from WebGPT. This was for GPT3. Since Wraith is now powered by GPT4 (more on that below), the intrinsic hallucinations should be fewer, and using internal web browsing to find good information should further enhance truthfulness and accuracy of output.

Updates

More Robust AI Content Evasion

Even before this update, Wraith can easily defeat GPTZero and many other AI content detectors. Now, it can consistently defeat copyleaks as well.

Caveats and challenges

Informal tone can consistently defeat copyleaks, but new "academic" tone may not. This is likely because academic / professional tones are more inherently robotic-sounding in nature, and a lot of these AI content detectors are trained on more academic sources. That said, academic tone'd article should still be able to defeat Copyleaks (and definitely other AI content detectors) most of the time. AI evasion challenges are 3-fold:

1. It used to be that higher temperature + penalties can help evade detectors effectively; they still do. But not Copyleaks since they've been in the game for many years. Completely random (or highly random) text is marked as AI-generated, as is text with burstiness / perplexity that's too low. Thus, challenge is having AI be 'creative enough' to have high burstiness and perplexity, but not so creative that Copyleaks detects it as AI-generated. Turns out: it's a very narrow band of creativity for it to pass Copyleaks.
2. GPT4 (and all GPT) products are getting dumber and dumber by the week. This means basic prompts that were working perfectly are now no longer being adhered to. This makes having GPT write human text harder and harder. A possible solution is to fine-tune our own LLM model to write humanlike text, but that requires capital that Wraith does not currently have. Thus, the more users we have, the better Wraith becomes. If you want to help out, and make money, we've got an affiliate program that pays out 40%, forever.
3. Copyleaks is getting stricter and stricter. At the beginning of me working on this upgrade (just last month), prompts were passing Copyleaks at 99.9% human. As were my actual text that I'd type in. Now, even my human text is classified as only 50-70% human a lot of times. So Copyleaks is skewing more and more towards false positives whereas GPT is skewing more and more towards not following instructions. For example, this article is written completely by me since I like writing these updates, but look at the results that Copyleaks gives for this article (see below).

Copyleaks strictness is further observed if you look at some competitors that claim to evade Copyleaks and look at their samples--just a few weeks ago they had about 60% human but now their samples are flagged as completely AI. Still, despite all these, Wraith still pumps out human written text for Copyleaks (at about the same % as if I were to type words in myself).

Bug Fixes

An issue was the server might crash from time to time due to lack of resources. Not as much a bug but a resource issue. I've since doubled the memory resources for the server and QoS should thus be a lot better. The algorithm to calculate the inherent meanings of connected blogs / external web search was also updated to consume less memory so it should both be faster and more efficient.

Better Engagement

I spent a lot of $$$ trying and retrying various prompts so GPT4 can one-shot a somewhat engaging prose for readers. This should hopefully ROI for you in the long-run, when coupled with AI content evasion. This has a lot more work to go because the 'perfect prose' is hard to achieve, but there are features coming later that'll let you customize tones and have you "deepfake" your writing style. Not only is the tones finely tuned to make it more engaging to read though, we've also included tables and charts.

Tables & Charts

Tables are just a way to condense information so it's easier for users to skim a section at-a-glance and know what's going on.

But flowcharts generated help summarize important parts of the article and looks something like this:

It is more useful for articles where sections are intended to be well-connected and ideas have some cause-and-effect to them though. It may not be optimal to use it for a listicle that lists out 17 places to visit since there's no logical flow between place #1 vs #2 vs #3.

Intro Enhancements & FAQ

Further, intros are now enhanced with bulleted points of key takeaways in the article, as SEO seems to like that. Additionally, articles come with an FAQ section at the end, again, for better SEO. Coming soon: dynamic FAQ schemas so that you can edit the FAQs and a dynamic FAQ schema will be generated. FAQ schemas are basically just some code to hint to Google that this article answers some questions so that not only will you have a chance to rank for your primary keyword for that article, but you'll have a chance to rank for any of the questions created in the FAQ. 1 article, many opportunities. Not only that FAQ schemas will be useful since it'll allow your ranked article to take up more search engine real estate. But more on that later when it's done being implemented.

More Intentful H2 Headings

GPT will now detect the type of article you're trying to write more closely (whether it's instructional, or a listicle, etc)--and will come up with an H2 structure that's most appropriate with your article's intent. How-to articles for example will generally have H2s that are "Step 1: Do This", "Step 2: Do This", etc. By aligning H2 titles with article and search intent, you should be able to get more optimal SEO results since it'll signal to scrapers that your article intent is aligned with the user's search intent. This gives you a leg up against competitors / amateurs that do not optimally structure or name their H2.

Fact-based GPT without plagiarism

When external links is enabled, Wraith will do a bunch of web searches for you and visit all the websites and attempt to understand the meaning of every single sentence of those websites. Then, Wraith will proceed to come up with ideas / write the articles for you. This way, the articles are going to be more accurate, grounded in reality, and also more in line with content that ranks #1 across various, related search terms, but all compiled into 1 article. Bonus: When turning this on, external links are included and sourced--but when scanning the text, there's no plagiarism detected. This means if you're blackhat (though we do not recommend this), you can remove attribution and it will still be considered original content. We do not recommend this, however, as having external links isn't a bad thing especially when giving appropriate credit (probably better for SEO even); whereas plagiarism, in the long run, can have a negative business impact.

There's only a few AI products that use external sources to assist in the writing process. But scanning their corpus generally yields a very high level of plagiarism.

Wraith passes plagiarism checks with flying colors. See below for an article that:

1. Does research on various places to visit in Europe
2. Uses that info to write

The images and external links are excluded in their scan when I copy/paste it, but here's proof that Wraith did include the links for you:

Conclusion

From most up-to-date AI evasion techniques, to better writing styles, to web based search to flowchart and table generation and little nuggets of SEO goodies, Wraith is now significantly better than its last version.

The best part? Tons more insane updates to come that'll make it much, much better than this version.