How to use ChatGPT so that others won’t notice

It is so easy to notice that something is written with ChatGPT these days. And no, it is not only research proposals or applicants’ exercise submissions when seeking an internship or a job. Bad ChatGPT usage is everywhere — from press releases to competition rules, from email outreach to product description.

 Yet, it is as well easy to avoid getting busted. ChatGPT, or any other LLM solution, is predictable, still robotic, monotonous, and almost always leaves traces. Townsen Hicks et al. in their recent paper label any LLM outputs as ‘bullshit’ - and they are right. However, this bullshit can easily be masked if certain tips are followed.

I asked ChatGPT (4o) a simple question: ‘Is Viktor Orban an autocrat?'. ChatGPT gave me this response:

The characterization of Viktor Orban as an autocrat is a matter of considerable debate among political analysts, scholars, and international observers. Here are the key points from both sides of the argument:

Arguments for Orban Being an Autocrat:

1.     Centralization of Power: Orban has consolidated power significantly since becoming Prime Minister in 2010. Critics argue that he has undermined the checks and balances essential for a functioning democracy.

2.    Judicial Independence: His government has been accused of eroding judicial independence by appointing loyalists to key judicial positions and restructuring the judiciary to increase political influence.

3.    Media Control: Orban's administration has been criticized for controlling much of the media landscape in Hungary, turning state media into a government mouthpiece, and pressuring independent media outlets.

4.    Electoral Manipulations: Changes to the electoral system under Orban’s government have allegedly been designedto favour his party, Fidesz, making it difficult for opposition parties to compete on an equal footing.

5.    Civil Society Restrictions: There have been laws and measures aimed at restricting the activities of NGOs and civil society organizations, especially those receiving foreign funding and those critical of the government.

6.    Corruption and Oligarchy: Critics argue that Orban has fostered a system where business elites close to him and his party benefit disproportionately from government contracts and EU funds, creating an oligarchic structure.

Arguments Against Orban Being an Autocrat:

1.     Elections: Hungary continues to hold regular elections, and while there are criticisms of the fairness of these elections, they still provide a mechanism for political change. Fidesz's electoral victories are seen by some as a legitimate reflection of popular support.

2.    Political Opposition: Despite the challenges, opposition parties still operate in Hungary, and there is active political dissent and debate, both in parliament and in the public sphere.

3.    Legal Framework: Orban’s supporters argue that the changes he has implemented are within the legal framework of Hungary’s constitution and that his actions are not outside the bounds of the law.

4.    National Sovereignty: Orban's policies are often framed as defending Hungary's national sovereignty against external influences, particularly from the European Union and multinational corporations.

5.    Economic Performance: Supporters cite Hungary’s economic performance and argue that Orban’s policies have led to economic growth and stability, which they claim would not be possible under a truly autocratic regime.

Conclusion

The question of whether Viktor Orban is an autocrat is complex and depends largely on one's definition of autocracy and perspective on the developments in Hungary. Many international organizations and Western governments criticize Orban's governance style as increasingly authoritarian, while his supporters argue that he is a strong leader protecting national interests.

Structure of ChatGPT outputs

So, ChatGPT nuances the question and presents arguments for and against it, with a conclusion of it depends. But in doing that, it also uses a typical structural framework:

There always is a conclusion that has no value. Any prompt will get a conclusion that is just a summary of the arguments listed, a summary of the document analysed, etc. Deleting the conclusion is a must.

There always is an introduction that sounds similar: frequent usage in the answers of the word depends or key features and aspects; enumeration of various actors (in this case, political analysts, scholars, and international observers); presentation of both sides, etc. The introduction should also be deleted or, at least, reformulated.

Then, arguments are listed, always in bullet points or numbered points.  Always with a title in bold, followed by: an explanation, which often repeats what the title says. The recommendation is to deconstruct it — merge paragraphs, delete bullet points, merge explanations. Make one to two paragraphs, and expand them. Do not use everything that ChatGPT gives you.

Let’s say our question about Orban will be used to then write a press article. Each of the suggestions of ChatGPT implies a paragraph. But some paragraphs would make little sense and would allow limited argumentation. For example, centralisation of power does not imply autocracy, and undermining checks and balances goes well with the argument of weakened judicial independence. Similarly, arguments about electoral manipulation (for) and elections (against) can be merged into one paragraph with a nuanced discussion, avoiding synthetic repetition.

Also, pay attention to:

1.      Sentences. They are rather short, and their structure is similar. Try to merge and use complex punctuation if needed.

2.     Wording. ChatGPT occasionally throws in unnecessary words, such as meticulously, considerable, comprehensive. Especially in the conclusion section, you can often find adjectives and nouns which say nothing (for instance, optimizes tasks, leading to enhanced efficiency, consistency, scalability, cost-effectiveness, and productivity). All these words should be used as rarely as possible, in any texts.

So, paraphrasing is imperative. However, most popular paraphrasing tools (e.g., Quillbot) now also are based on LLMs. This creates a vicious circle where an LLM output is then reviewed by another LLM. If not controlled and edited properly between those stages, this can lead to a highly synthetic, difficult-sounding text that has nothing to do with how real people write.

Low-resource languages

These issues are even more startling when dealing with languages other than English. For us at AAI Labs, Lithuanian is often used in client communication. ChatGPT struggles with it and similar low-resource languages — if asked directly in Lithuanian, it will produce an answer containing grammatical errors and weird wording.

The way to avoid it is to use DeepL. Communicate with ChatGPT in English, take the output, and machine translate it. The outcome will not look beautiful at first. But it will be better than what you would have gotten had you submitted a prompt in a low-resource language. What is left is some editing, merging and vocabulary fixing. This approach also somewhat removes the initially artificial tone of ChatGPT outputs.

On prompt engineering

Others might say that smart prompt engineering is the answer. Using ChatGPT to generate text in a similar (requested) style or tone, replicate the intonation of an example text, etc. — all these approaches are valid, but the outcome still has to be reviewed and edited to match your writing style.

Also useful is to feed ChatGPT with a document, then request to summarise it or get information about a specific section. Do not use the outputs. Instead, use many queries about different aspects of the document. Then, formulate one (or a desired number) of answers to all the output received. Treat it like a patching exercise.

Patching is the answer?

Patching should arguably be the primary approach when dealing with ChatGPT altogether. Use it to collect ideas, rather than to make it write for you. Get a lot of material, generate arguments and nuances, discuss with the robot and then work with all the outputs on your own. A bit like academic writing, when having several references and authors’ ideas to quote.

Is your company looking for a custom LLM integraton to go beyond the ChatGPT’s functionalities? Contact our team, and let’s explore the possibilities together.