AI and credibility: how to avoid plagiarism and hallucinations
The rapid deployment of major language models in our productivity tools promises to accelerate the writing, translation, and structuring of knowledge. However, behind the promise of a "publish-ready" text lie risks that can undermine the credibility of an author, a research team, or a company: unsourced reproduction of pre-existing ideas, inventions presented as facts, and phantom bibliographic references. In other words, technological innovation never replaces constant intellectual and ethical rigor.
Three swords of Damocles hang over every user of generative AI. The first is the plagiarismSometimes unintentional: a model might reproduce word for word – or almost – a passage it has seen millions of times, without indicating the source. Even more subtle, it might regurgitate a text that has already been plagiarized, creating a "double plagiarism" that spreads like a chain reaction of misunderstandings. The second is the hallucinations : factually false but linguistically plausible statements that no one detects on the first reading. Finally, the third concerns the manufactured references : bibliographic records invented from scratch or accompanied by non-existent identifiers (DOI, PMID).
Recent analyses reveal the scale of the problem: exactly half of the references generated by chatbots turn out to be entirely fictitious; nearly four out of ten DOIs are incorrect; and only 7% of automatically generated bibliographic references are perfectly authentic. Extrapolating to the technical literature as a whole, these percentages translate into thousands of misleading citations that then contaminate internal reports, blog posts, and client presentations.
To get out of this turbulent zone, the authors of the document propose a three-level usage map.
Tier 1 This corresponds to safe practices: spelling correction, stylistic smoothing, grammatical adjustments, or translation of a text already written by the author. In these cases, AI assumes the role of an editorial assistant; no new ideas are introduced, and the risk of factual error remains limited, provided a quick proofreading is performed.
Tier 2 This refers to a cautious use: expanding a section from raw notes, generating a detailed outline, summarizing a text you have written, or guided brainstorming. AI then becomes a thinking partner, but the user must verify each argument, rephrase to preserve their voice, and above all, compare the suggestions to the company's strategic context.
Tier 3 This coincides with the red zone: writing entirely new content, automated literature reviews, preliminary data analysis, or generating a complete bibliography. Here, automation bypasses in-depth understanding and multiplies the opportunities for error. If in doubt, it's best to abandon the process and revert to a more manual approach, or to confine these experiments to a sandbox environment, far removed from official publication.
Before any external dissemination – scientific article, white paper, strategic memo or NextStart.ai blog post – a simple four-question filter helps to define the scope:
- Are my ideas still truly my own?
- Have I maintained my critical skills rather than delegating most of them?
- Has each fact, figure, and reference been verified in a reliable primary source?
- Is the contribution of AI described honestly and accurately?
If a negative response appears, the manuscript must be revised. This requirement also aligns with the consensus of international publishing organizations: total transparency regarding the use of automated tools; the impossibility for AI to sign or assume responsibility for a text; and the primacy of human expertise for making decisions, validating, and contextualizing.
It is therefore necessary to establish a internal charter describing the permitted, monitored, or prohibited scenarios. Then, train every writer—whether they write code, documentation, or marketing materials—to detect and correct unbelievable output. Implement a verification pipeline The model drafts the text, but a search engine checks each quote, and a human completes the process by judging its relevance. Finally, the added value is measured: is a clearer text also more accurate? Is a time saving worth a loss of understanding? Answering these questions fosters a product culture focused on reliability, not just the "wow" factor.
A good reflex is to create a Internal repository of validated sources Articles, public statistics, and in-house reports are used. Before asking a model to generate a paragraph, it is forced to draw from this database rather than relying on guesswork. Drafts are then scrutinized by similarity detection tools and by reviewers responsible for ensuring scientific integrity. This double safeguard—algorithmic control and human oversight—reduces the risk of factual error and the suspicion of plagiarism.
The goal is not to stifle innovation, but to integrate it into a sustainable cycle. Language technologies facilitate the democratization of knowledge and creativity, provided they are managed like a co-pilot They propose, we dispose. Let's ask them to reformulate, structure, visualize, but let's reserve for ourselves the interpretation, the nuance, the final responsibility.
At a time when algorithms produce more words than they read, scarcity is shifting towards critical thinking. A truly distinctive text is not merely fluent: it demonstrates a personal grasp of ideas, a critical examination of data, and an awareness of its own limitations. It is this ability to combine speed and vigilance that will distinguish interchangeable content from that which permanently establishes the author's reputation.
The future will belong to those who combine the generative power of AI with a ethics of attention Pay attention to the truth, pay attention to the other, pay attention to the consequences. This stance sometimes requires slowing down to better sort, simplifying to better verify, and questioning the apparent obviousness of a text that is too perfect to be honest.
Ultimately, generative AI doesn't impose a revolution in ethics; it simply brings it into the open. The classic principles of research—originality, accuracy, traceability—remain the guiding compass. Their needles only spin wildly if human vigilance is switched off. Let's therefore keep a firm grip on this compass: it's what ensures that the path of innovation leads to that of trust.
To go further: this requirement for rigor is reflected in our analysis on chatbots that blur the lines between truth and falsehood and our guide to mastering the art of prompting and guiding AI.