Harness the Lionbridge Lainguage Cloud to support your end-to-end localization and content lifecycle

Lainguage Cloud™ Platform
Connectivity
Translation Community
Workflow Technology
Smairt Content™
Smairt MT™
Smairt Data™
Language Quality
Analytics

Our people are our pride, helping companies resonate with their customers for 20+ years. 

About Us
Key Facts
Leadership
Insights
News
Trust Center

 

SELECT LANGUAGE:

person with streams of data and light

AI Self-Correction

What to Know Now

Embracing Generative AI is crucial to succeeding, especially when your competitors are doing the same for their workflows, translation, or content creation and optimization. One critical element of using Generative AI is self-correction. Unfortunately, Large Language Models (LLMs) can deliver output with inaccuracies due to a few factors. This is because the data an LLM is trained on might include problematic or wrong information. AI tools will also sometimes “hallucinate,” or make up information. To address issues in AI output, it’s possible to take “self-correction” measures into the initial set of prompts. (Some experts have also called it “self-critique” or “self-refine.) Multiple studies have tested methods that require LLMs to review their output and refine responses before delivering them. Read our blog post to learn some of the techniques people are using to implement self-correction in their AI solutions (or having their AI solutions provider do it). This blog will also cover the limitations of AI self-correction.

Common AI Self-Correction Tactics

These are four ways people are currently implementing AI self-correction:

1. An accuracy-focused prompt: Sometimes, adding a prompt emphasizing accuracy into the series of prompts is effective. Here is a popular one posted on X:

“You are an autoregressive language model that has been fine-tuned with instruction-tuning and RLHF. You carefully provide accurate, factual, thoughtful, nuanced answers, and are brilliant at reasoning. If you think there might not be a correct answer, you say so.”

2. Turning AI tools into an expert: One way to preempt inaccuracies is to turn your AI tool into an expert less likely to make mistakes. Many users and AI service providers, including a group of GitHub developers, are creating prompts that command AI tools to act like experts. Notably, the best expert personas are the ones with the most detail about following best practices — provided they’re widely accepted. With commands that are too general, the AI tool may begin to hallucinate its content. For example, saying, “You’re an excellent career counselor,” isn’t enough. The prompts should include guidance for best practices that career counselors generally follow. Another best practice is to test the series of prompts with a task you know the answer to. This will help determine where to optimize the expert persona prompts. Sometimes, it may even make sense to develop multiple iterations of an expert persona prompt depending on the type of task. The GitHub developers made a list of 15 series of prompts they used to turn AI into an expert assistant. Though they aren’t the only ones to do this, their list is notably comprehensive.

AI Expert Prompts:

1. Career Counselor

2. Interviewer for a specific position

3. English Pronunciation Helper

4. Advertiser

5. Social Media Manager

6. AI Writing Tutor for Students

7. Accountant

8. Web Design Consultant

9. Act as a UX/UI Developer

10. IT Architect

11. Cyber Security Specialist

12. Machine Learning Engineer

13. IT Expert

14. Generator of Excel formulas

15. Personal Chef

3. Adding “pre-hoc” or “post-hoc” prompting: It’s possible to add prompts that modify the style of AI output. Perhaps the style needs to be more formal or informal, or targeted towards highly-educated audiences or audiences with a high school-level education. If the prompts are added after the output is generated, this is called a “post-hoc prompt.” Per a recent research project from Google’s DeepMind, the best results occur with equally strong pre-hoc and post-hoc prompting.

orange folds with AI image

4. Using prompts to address biases: If LLMs aren’t trained on the right data, their output may reflect the biases of the millions of people who spew hateful content on the Internet. Per a recent study by the Anthropic AI lab, it’s possible to use Reinforcement Learning from Human Feedback (RLHF) to train an LLM to produce output without (or with less) racism, ageism, misogyny, etc. Include instructions in the AI’s constitution to consider general ethical principles your team decides upon. Part of this process is adding a line into prompts that preempts the LLM from relying on harmful stereotypes or philosophies. In some cases, AI tools have been shown to begin “positively discriminating” in their output, which may even exceed expectations.

Limitations of AI Self-Correction

While AI self-correction measures may be powerful, studies also show that it still has limitations. The same Google DeepMind study found that LLMs sometimes actually perform worse with self-correction measures. In cases where it doesn’t impair performance, self-correction isn’t consistently effective for every series of AI prompts, especially where external sources (a calculator, code executor, knowledge base, etc.) aren’t used. For best results, self-correction measures need access to a benchmarked data set with basic truths built in. With these references, the AI tool will know when to stop its reasoning process, thus avoiding overcorrecting its output. Of course, the researchers noted that some tasks are too complex to enable providing an AI tool with these kinds of references.

The same study also found that another limitation of AI self-correction occurs when multi-agent LLM applications are used. The LLM is asked to perform multiple tasks as different “agents,” or actors.

Examples:

The LLM generates code as one agent. Then, it also checks the code as another agent.

The LLM performs a debate, with one agent taking each side.

The problem occurs because the multiple agents use a form of majority voting to decide which answer is correct, thus creating a kind of echo chamber or “self-consistency,” rather than true accuracy.

series of buttons and charts

The Value of a Human in the Loop

The limitations of AI self-correction underscore how essential a human in the loop is. AI tools can enhance translation efficiency, but they often need human intervention at some point. Perhaps a human must develop the best series of prompts, check an initial sample, or review output at the end. Self-correction measures may assist with the entire process, but can’t replace a human in the loop.

To that end, it’s vital to work with AI consulting experts, like the ones at Lionbridge, who can help address the AI trust gap. They should:

  • Minimize the risk of untrustworthy or low-quality content/output

  • Ensure security of data from cyber-attacks or any kind of compromise

  • Be creative and help develop new, engaging, and original content or output

  • Check and correct for accuracy, especially for complex material that requires intensive education or expertise

  • Never try to sell you unnecessary technology, solutions, or subscriptions

  • Share the entire process and invite input, feedback, and customization throughout

Get in touch

Interested in learning how to utilize AI to automate content creation, website content optimization, or other language services? Lionbridge’s dedicated team of AI experts is ready to help. Let’s get in touch.

linkedin sharing button

AUTHORED BY
Samantha Keefe and Thomas Gaitley, Director of Resiliency
  • #technology
  • #translation_localization
  • #ai
  • #blog_posts