{"id":13134,"date":"2025-01-16T10:14:56","date_gmt":"2025-01-16T09:14:56","guid":{"rendered":"https:\/\/proetcon.de\/?page_id=13134"},"modified":"2025-11-10T13:00:31","modified_gmt":"2025-11-10T12:00:31","slug":"ai-tools","status":"publish","type":"page","link":"https:\/\/proetcon.de\/index.php\/en\/software-migration-2\/ai-tools\/","title":{"rendered":"AI Tools and Software Migration"},"content":{"rendered":"
\t\t\t

HOME<\/a> > SOFTWARE MIGRATION<\/a> > AI<\/a><\/strong><\/p>\n<\/div>\n\t\t<\/div><\/div><\/div><\/div>

\t\t\t

\"SoftwareSOFTWARE MIGRATION AND AI<\/h2>\n
\n

The British compu\u00adter scien\u00adtist Alan Turing is regarded as one of the most influ\u00aden\u00adtial pioneers in the field of artifi\u00adcial intel\u00adli\u00adgence (AI). In 1950, he formu\u00adla\u00adted an intel\u00adli\u00adgence test for compu\u00adters: a compu\u00adter is intel\u00adli\u00adgent if, in a conver\u00adsa\u00adtion with it, a person cannot recog\u00adnise whether it is a compu\u00adter or a human.<\/p>\n

Today, Turing\u2019s visions have become reality. By recog\u00adnis\u00ading statis\u00adti\u00adcal corre\u00adla\u00adti\u00adons between words, AI language models can deter\u00admine the most likely next word (seman\u00adtic proba\u00adbi\u00adlity) and enable human-like communication.<\/p>\n

The current areas of appli\u00adca\u00adtion for AI go far beyond commu\u00adni\u00adca\u00adtion. This raises the question: what are the possi\u00adbi\u00adli\u00adties and limits of AI in software migration?<\/p>\n

In the follo\u00adwing, various AI tools are analy\u00adsed with regard to their use in software migration.<\/p>\n<\/div>\n<\/div>\n\t\t<\/div><\/div><\/div><\/div><\/div><\/div>

\t\t\t

Defini\u00adtion of the subject matter<\/h3>\n

In the follo\u00adwing, \u201csoftware migration\u201d refers to tool-supported, compi\u00ad\u00adler-based software migration; speci\u00adfi\u00adcally, the process of conver\u00adting program\u00admes from a legacy language to a modern programming language using conver\u00adsion tools. Alter\u00adna\u00adtive migration approa\u00adches such as manual migration are not conside\u00adred. The support of migration projects by e.g. analy\u00adsis tools or tools for test support etc. are also not included in the follo\u00adwing approach.<\/p>\n

The inves\u00adti\u00adga\u00adti\u00adons include the conver\u00adsion of a typical COBOL programme to Java and its proces\u00adsing: on the one hand with various AI tools available on the net, on the other hand with the CoJaC (COBOL to Java Conver\u00adter) conver\u00adsion tool develo\u00adped by pro et con. The results are compared and evaluated.<\/p>\n

The evalua\u00adtion is inten\u00added to provide infor\u00adma\u00adtion on the extent to which the results of the AI tools are seman\u00adti\u00adcally equiva\u00adlent to the correct results provi\u00added by the COBOL compi\u00adler, for example. Seman\u00adtic equiva\u00adlence is an essen\u00adtial, if not the most important criter\u00adion in a migration project.<\/p>\n<\/div>\n\t\t<\/div><\/div><\/div><\/div><\/div><\/div>

\t\t\t

Test basis<\/h3>\n

The test is based on a simple COBOL programme. The programme has no logic. However, it conta\u00adins typical COBOL instruc\u00adtions, which are used in many produc\u00adtive environ\u00adments. Experi\u00aden\u00adced COBOL programm\u00aders will under\u00adstand the COBOL programme immedia\u00adtely. Diffe\u00adrent data defini\u00adti\u00adons are made with simul\u00adta\u00adneous initia\u00adli\u00adsa\u00adtion. The data is manipu\u00adla\u00adted, e.g. through value assign\u00adments or overlays, and repre\u00adsen\u00adted with DISPLAY after each manipulation.<\/p>\n

Micro\u00adFo\u00adcus COBOL Compiler:<\/strong><\/p>\n

The programme was compi\u00adled with the Micro\u00adFo\u00adcus COBOL compi\u00adler and then execu\u00adted. It delivered the follo\u00adwing results.<\/p>\n

<\/p>\n<\/div>\n\t\t<\/div><\/div><\/div><\/div>

\t\t\t

<\/p>\n<\/div>\n\t\t<\/div><\/div><\/div><\/div><\/div>

\t\t\t

AI TOOLS<\/h2>\n

The experi\u00adments were conduc\u00adted with the follo\u00adwing AI tools: ChatGPT (GPT-4o and GPT\u20113.5), DeepSeek, Copilot, CodeCon\u00advert, CodeGPT, Cursor and Black\u00adbox AI.
\nFrom the COBOL programme, Java program\u00admes, which should have identi\u00adcal function\u00ada\u00adlity, were genera\u00adted by all above mentio\u00adned AI tools. It should be noted that the various AI tools were not speci\u00adfi\u00adcally trained for the task at hand.<\/p>\n

The genera\u00adted Java program\u00admes, inclu\u00adding the results at runtime, are documen\u00adted below. The focus lies on the two AI tools ChatGPT-4o and DeepSeek. The tests of the other AI tools analy\u00adsed delivered identi\u00adcal results:<\/p>\n<\/div>\n\t\t<\/div><\/div><\/div><\/div><\/div><\/div>

\t\t\t

ChatGPT-4o<\/a><\/h2>\n

\"Java<\/a><\/p>\n

When the conver\u00adted programme is execu\u00adted, it returns incor\u00adrect results.
\nA second conver\u00adsion produ\u00adces a diffe\u00adrent source code.
\nThe (incor\u00adrect) results of both conver\u00adsi\u00adons differ.<\/p>\n

Learn more in our detailed analysis.<\/a><\/p>\n<\/div>\n\t\t<\/div><\/div><\/div>

\t\t\t

DeepSeek<\/a> (v3.1)<\/h2>\n

\"Java<\/a><\/p>\n

When the conver\u00adted programme is execu\u00adted, it returns incor\u00adrect results.
\nA second conver\u00adsion produ\u00adces a diffe\u00adrent source code.
\nThe (incor\u00adrect) results of both conver\u00adsi\u00adons differ.<\/p>\n

Learn more in our detailed analysis.<\/a><\/p>\n<\/div>\n\t\t<\/div><\/div><\/div><\/div><\/div>

\t\t\t

<\/p>\n<\/div>\n\t\t<\/div><\/div><\/div><\/div>

\t\t\t

COJAC (COBOL TO JAVA CONVERTER)<\/h2>\n

Besides other tools, CoJaC is used in pro et con migration projects for conver\u00adting COBOL program\u00admes to Java.
\nThus, compa\u00adring the results of CoJaC with those of the AI tools is appropriate:<\/p>\n

The COBOL programme was conver\u00adted to Java using the CoJaC tool from pro et con.
\nThe Java programme was compi\u00adled and execu\u00adted. At runtime, the Java programme delivered the correct results, identi\u00adcal to the COBOL programme.<\/p>\n

Logically, a new conver\u00adsion also delivered identi\u00adcal Java code and identi\u00adcal, correct results at runtime.<\/p>\n

<\/p>\n<\/div>\n\t\t<\/div><\/div><\/div><\/div><\/div>

\t\t\t

DISCUS\u00adSION OF THE FINDINGS<\/h2>\n
    \n
  1. All AI tools tested deliver incor\u00adrect results. The reason for this is that COBOL-speci\u00ad\u00adfic language elements are mapped using Java on-board resour\u00adces. This only works to a limited extent, as there are no seman\u00adti\u00adcally equiva\u00adlent Java data types\/variables for the complex COBOL data descrip\u00adti\u00adons and their proper\u00adties, for example. AI models make predic\u00adtions based on proba\u00adbi\u00adlity, which can only appro\u00adxi\u00admate 100% seman\u00adtic equiva\u00adlence. Although the under\u00adly\u00ading machine or deep learning models can be optimi\u00adsed through ongoing training, it remains to be seen whether \u2013 and if so, with what training effort \u2013 this 100% seman\u00adtic equiva\u00adlence can be fully achie\u00adved. In contrast, CoJaC has a library in which all data descrip\u00adti\u00adons in COBOL are included in the form of Java classes and packa\u00adges. These emulate the COBOL data descrip\u00adti\u00adons in the Java code. This ensures seman\u00adtic equiva\u00adlence between the COBOL and Java programme.<\/li>\n
  2. Upon repea\u00adted Java genera\u00adtion, all AI tools tested delivered a modified source code. Thus, a conver\u00adsion by an AI tool is not repli\u00adca\u00adble 1:1. In a migration project usually compri\u00adsing several million LOC, it is neces\u00adsary to repea\u00adtedly convert a certain number of program\u00admes (e.g. for further develo\u00adp\u00adments) in several itera\u00adti\u00adons. A source code that is constantly changing as a result creates discrepan\u00adcies in related compon\u00adents (referen\u00adcing, classes, inter\u00adfaces), as each programme commu\u00adni\u00adca\u00adtes via inter\u00adfaces with other program\u00admes in the programme system. This is unaccep\u00adta\u00adble in a migration project. With a compi\u00ad\u00adler-based migration tool such as CoJaC, a programme conver\u00adsion can be repea\u00adted as often as requi\u00adred and always delivers identi\u00adcal source code.<\/li>\n
  3. Quali\u00adfied COBOL programm\u00aders will have recog\u00adnised that the COBOL programme is incor\u00adrectly programmed at several positi\u00adons (e.g. a two-digit number is assigned to a one-digit varia\u00adble). This is obviously a questionable style, but indeed happens in reality, especi\u00adally in long-running, complex projects, which is usually the case with legacy software. Nobody can guaran\u00adtee that, at the time of a migration, every\u00adthing will be programmed correctly. And a migration must do justice to this fact by mapping every special case 1:1 seman\u00adti\u00adcally equiva\u00adlent for criti\u00adcal software. This is exactly what AI tools defini\u00adtely cannot do, whereas the compi\u00ad\u00adler-based approach can.<\/li>\n
  4. \u2018The\u2019 COBOL does not exist. The language depends on the hardware platform, compi\u00adler dialects and project-speci\u00ad\u00adfic compi\u00adler settings, which change the behaviour of the system and are not repre\u00adsen\u00adted in the COBOL code itself. A wide variety of precom\u00adpi\u00adlers are common as well. This means that every legacy system is indivi\u00addual and requi\u00adres equally indivi\u00addua\u00adli\u00adsed proces\u00adsing. This is another area where AI tools reach their limits, whereas the compi\u00ad\u00adler-based approach supports this custo\u00admi\u00adsed processing.<\/li>\n<\/ol>\n<\/div>\n\t\t<\/div><\/div><\/div><\/div><\/div><\/div>
    \t\t\t

    CONCLU\u00adSION<\/h2>\n

    <\/p>\n