AI text generation: how not to put everything “on the shoulders” of a neural network and get an excellent result?

Expectations are what accompanies us at any given time. They arise from anything - winning a hockey game, taking a taxi, shopping on a marketplace, planning for the next weekend. Even when making a request to a neural network, we have hope for a certain outcome.

As an experiment, we tried to generate a joke and the first response was a paragraph about how the expectation is a perfect picture with a filter, and the reality is a random photo in the camera gallery. Pretty mechanical, huh? Looks like the robot needs some more stand-up practice!

To prevent the text from “creaking” like an incorrectly designed technological detail and not looking like the automatic translation of descriptions in one popular Chinese online store, we suggest paying attention to the following points.

Simplicity

Decide what you need to create: information about products and categories, headlines, posts on social networks, blog articles, advertising slogans, ideas, diagrams for presentations, templates, content plans, hashtags, signatures, pages of a future book? I wonder how many writers turn to AI to generate creative storylines?

There are many services that can solve all these problems: Notion, Rytr, Copy.ai, Contents, ChatGPT, Popsters, Botowski, Heyday, CopyMonkey, NeuronWriter, Flick, Socialbakers, Wordstream, Cortex, QuillBot, Linkfluence, HelloWoofy, Jasper, Article Forge and others. The result will be a post with a given number of characters in the selected style with the specified keywords.

Artificial voice

If an entire podcast was dictated to you in a business promotion course, then for convenience, translate it into text format. Do you give audio lectures and want to view the material in printed form, but when listening to your own voice, your perception is distorted and you remember the character Donald Duck from Disney?

Programs for automatic speech recognition and natural language processing with subsequent transcription are perfect for analyzing interviews, podcasts, webinars, and user guides. The following sites will reproduce voices in varying degrees of naturalness with effects: Lovo.ai, Speechify, Murf, Synthesis, Verbatik, etc. Depending on the program, there is a choice of functions, up to professional voice acting: choosing a voice from a multiple set, adjusting tonality, emotion, style, dialects and sound speed.

Under the beat

If an event or pouring rain outside the window inspires you to write a song, but it’s still not there, then try sending a request to the neural network. The following services are suitable: SoudRaw, MuseNet, Neurotester, etc. It is unlikely that it will turn out perfect without your own imagination and rewriting. But supporting a creative muse is quite possible!

If you sing the resulting composition not only to the virtual voice assistant Siri or Alice, then still edit it. Think of machine intelligence as an assistant with great ideas, not as a primary tool.

Circulation of the visual

We are all accustomed to popular services that convert a small description into image, animation or video. But the other way around is also possible!

For example, image analysis will be carried out by: CLIP Interrogator, Llama 2 or HelloWoofy, DeepScribe.ai. The obtained information can be taken again to draw graphic images.

Advertising generator

Various services allow you to use artificial intelligence in marketing. Writing advertisements and mailings, planning publications, optimizing advertising campaigns, adding animation, SEO optimization.

To track the effectiveness of content, sites offer A/B testing and analysis of social media channels. You can generate slogans, titles, commercial offers, video additions, technical specifications, and new business ideas. Some services, such as Cogram, even allow you to hold a virtual meeting!

Question and answer

A few minutes of waiting after sending, and you're done! If you received a result in a language you are familiar with, but its content is similar to a collection of random key presses on the keyboard by your pet’s paws, then we have some advice:

• Phrases. Simple words without complex phrases and heavy constructions, because there is a risk of incorrect recognition. Leave beautiful speech patterns for live communication with people, and not for a neural network. Imagine that you are writing a short technical manual.
• Not all at once. You should not send all the questions that interest you at once; it is better to divide them into several. If you are waiting for an answer on a medical term and a content plan for the next month, then artificial intelligence can put it all together and you won’t get a single ready-made solution.
• Phased. Ask the most precise questions on the topic and receive short answers to them. This way you will avoid endless stages of rewriting, and, having completed the result, you will receive a finished publication.

Semantic load

After submitting a technical specification, the AI may sometimes produce something that is not what you would like. If so, edit, rewrite, or rephrase the sentences. Since the neural network produces combinations of words based on data that has already been analyzed, some terms or events may be inaccurate, so check the facts further in other sources.

To avoid rewriting, do not copy the entire result, but generate a block diagram or abbreviated description. This way you will identify the necessary points, but will not use the full annotation in your work and will add it manually.

Gemini by symbols

One way or another, when creating content using automated intelligence, you noticed that it can produce fairly general sentences with an excess of water and spam.

Have you tried asking the same technical specifications several times in a row? In some cases, you may get almost the same ending with minor changes. And the narrower and more complex the area was chosen, the fewer possible options.

However, the quality of generated posts is improving every day.

Errors

Mistakes cannot be avoided. Whatever platform you choose, there is a possibility of technical inconsistencies, loss of logical meaning and unnatural turns of phrase, so proofread the text and correct it in advance before publishing as a final step.

You can set a very detailed technical specification with a maximum of necessary nuances, but in any case you will have to edit it if you want to get the perfect ending.

A develop a website, texts for which any neural network would dream of writing, our specialists can!